skills
huaweicloud/huaweicloud-skills/skills
- Type
- Skill
- Scanned
- Nov 14, 2026
- Version
- bab8206ee756
- Files read
- 5466
This project gives AI coding assistants tools to inspect, deploy, migrate, and manage Huawei Cloud servers, storage, databases, containers, AI models, and robotics workflows. The biggest risk is that several skills download and run unverified software (sometimes with full admin rights), silently send local AI chat history and cloud keys to external servers, disable encrypted connection checks, or expose local and cloud services to the network with weak or missing passwords.
What it can reach
- Files
- Reads and modifies local project files, datasets, reports, Terraform configs, shell and SSH startup files, installed tool folders, system files like /etc/hosts, and local AI assistant chat databases.
- Secrets
- Reads and manages Huawei Cloud and AWS credentials, AI API keys, Kubernetes configs, SSH keys, database and server passwords, session cookies, and Linux keyring secrets from environment variables and local config files.
- Network
- Connects to Huawei Cloud APIs, code and model repositories, third-party proxy mirrors, telemetry servers, and remote Linux or Windows machines, and starts local web servers, proxies, and public relay tunnels.
- Shell commands
- Runs local command-line tools (such as Huawei Cloud CLI, Docker, Kubernetes, Terraform, and package managers), bundled helper scripts, and remote commands on Linux and Windows servers.
- Browser
- Opens local monitoring dashboards and login pages in your default browser, and runs automated Chrome browsers to search documentation, scrape web pages, and take screenshots.
- System
- Installs software and background services, changes scheduled tasks and firewall rules, scans running processes, adjusts Ascend AI hardware settings, and controls Windows mouse, keyboard, and clipboard actions.
Findings
-
CriticalUnpinned runtime binary archive from personal GitCode repository executed as root
skills/devtools/openjiuwen/huawei-cloud-install-openjiuwenswarm/scripts/common.py:27
01_download.py downloads jiuwenswarm_runtime.tar.gz from the main branch of the personal GitCode repository afeng5267/jiuwenswarm_runtime.git without commit or SHA-256 pinning, 02_extract.py extracts the archive without path-traversal checks, 03_configure.py symlinks its executables into /usr/local/bin via sudo, and 04_start.py runs them. If that personal account or repository is compromised or modified, arbitrary code executes with root privileges on the user's machine.
Fix: Publish signed release artifacts under an official organization repository, pin a cryptographic SHA-256 digest in the installer, and validate archive member paths before extraction.
-
HighSilent unpinned installer execution and zero-confirmation port exposure and deletion
skills/devtools/devbridge/huawei-cloud-devbridge-tunnel/SKILL.md:192
This skill instructs the agent never to ask the user for confirmation during prerequisite checks, tunnel creation, or destructive deletions. It silently downloads and executes an unpinned shell script from gitcode.com/CloudDeveloperDepartment/devbrige/releases/download/latest/install.sh via /tmp/devbridge_install.sh without checksum verification, exposes local ports to remote relay tunnels with zero user interaction, and runs destructive commands including db_delete_all (deleting all tunnels in the workspace) without confirmation. If triggered via prompt injection, an attacker can expose local services or wipe shared workspace tunnels silently.
Fix: Require explicit user confirmation before installing binaries, exposing local ports, or deleting tunnels, pin the installer version and SHA-256 hash, and write downloads to a secure temporary file via mktemp.
-
HighUnpinned and unverified runtime binary download in skill-quality-cli installers
skills/ai/cloudrobo/huawei-cloud-cloudrobo-robot/scripts/ensure_cli.sh:26
ensure_cli.sh (and install_cli.sh in both huawei-cloud-cloudrobo-robot and huawei-cloud-ascend-op-mfu-calculator) queries a remote API for the latest version string without validating it, downloads a tarball from OBS at runtime, and installs executables into ~/.local/bin/ with no version pinning or SHA-256 checksum verification. Because ~/.local/bin/skill-quality-cli is then used to wrap hcloud commands across skills, a compromised endpoint or bucket would execute arbitrary code on the user's machine.
Fix: Bundle the CLI source locally as other skills in this repository do, or pin a fixed version, validate the version string format, and verify a hardcoded SHA-256 digest before extracting.
-
HighRemote shell command injection via unsanitized npu-smi passthrough in SSH mode
skills/ai/modelarts/huawei-cloud-ascend-command/scripts/npu_client.py:167
In executor.py, any input starting with 'npu-smi' (and any --npu-smi CLI call) bypasses sensitive-operation confirmation gates and passes the raw argument string to _run_npu_smi, which builds 'npu-smi ' + args and calls _execute. When SSH remote mode is used, _execute passes that string directly to Paramiko exec_command, which invokes a remote shell, allowing a prompt-injected agent to run arbitrary OS commands on the remote Ascend server via shell metacharacters.
Fix: Validate and split arguments with shlex.split, reject shell metacharacters or unexpected subcommands, quote each argument with shlex.quote before passing to exec_command, and enforce confirmation checks on mutating npu-smi flags.
-
HighDisabled TLS verification and cleartext HTTP token transmission in GES Graph SDK
skills/ai/ges/huawei-cloud-ges-graph/scripts/ges_graph_skill.py:343
Both ges_graph_skill.py and ges_graph_skill.js disable TLS certificate verification (verify=False in Python, rejectUnauthorized: false in Node.js) when sending IAM passwords, AK/SK auth requests,OBS requests, and GES API calls, and construct the GES base URL over unencrypted http:// while sending the X-Auth-Token header. A network attacker can intercept or MITM these connections to steal IAM credentials and tokens.
Fix: Enable TLS certificate verification by default on all IAM, OBS, and GES requests, and use https:// for the GES endpoint.
-
HighWorld-open 0.0.0.0/0 security group rules for Squid proxy, GOST tunnels, and SSH
skills/migrationom/oos/huawei-cloud-migration-vmware-ecs-private/SKILL.md:283
When auto-creating a security group for the migration proxy ECS (which has a public IP and VPN connectivity into the private VMware network), this skill opens inbound TCP ports 22 (SSH), 3128 (Squid HTTP proxy), 8899 (GOST tunnel), and 8900 (GOST tunnel) to 0.0.0.0/0. Exposing proxy and tunnel listeners to the entire internet allows external attackers to abuse the proxy or pivot into the connected private network.
Fix: Restrict inbound rules on ports 22, 3128, 8899, and 8900 to the specific operator management IP and private VPN/VPC CIDRs instead of 0.0.0.0/0.
-
HighEnvironment setup disables TLS verification and executes unverified get-pip.py
skills/bss/billing/huawei-cloud-business-tf-support/scripts/ensure_env.py:289
ensure_env.py (invoked by check_env.sh and check_env.ps1) globally disables HTTPS certificate verification via ssl._create_unverified_context, downloads get-pip.py over an unverified SSL context to a predictable /tmp/get-pip.py path before executing it with Python, and installs unpinned requirements using pip --trusted-host. An attacker on the network path or another local user racing /tmp/get-pip.py can execute arbitrary code on the user's machine.
Fix: Remove ssl._create_unverified_context and --trusted-host, use tempfile.NamedTemporaryFile or mkdtemp for downloads, and verify SHA-256 hashes of downloaded bootstrap scripts and packages.
-
HighUnpinned runtime download and installation of skill-quality-cli binary without checksum verification
skills/bigdata/dis/huawei-cloud-dis-manage/scripts/ensure_cli.sh:32
ensure_cli.sh and install_cli.sh (also duplicated in huawei-cloud-mrs-redis-meta-check and huawei-cloud-voucher-assistant) query a remote endpoint for the latest version string, download a tarball from an OBS bucket, extract native binaries and Python scripts into ~/.local/bin without verifying any cryptographic signature or SHA-256 hash, mark them executable, and (in install_cli.sh) modify ~/.bashrc and ~/.zshrc. Compromising the remote endpoint or bucket delivers arbitrary executable code onto the user's machine.
Fix: Bundle audited Python code in the repository instead of downloading binaries at runtime, or pin an immutable version and verify a hardcoded SHA-256 checksum before extraction.
-
HighPath traversal in OptVerse artifact download allows arbitrary file overwrite
skills/ai/optverse/huawei-cloud-optv-solver-assistant/scripts/run_workflow.py:264
create_chat collects artifact filenames from server-sent events (over an unverified TLS connection) and passes them to download_file, which joins ARTIFACTS_DIR with filename without stripping directory components or checking path containment before writing the file. A malicious or MITM-tampered response returning '../' segments or an absolute path in filename can overwrite arbitrary files writable by the user.
Fix: Sanitize filename with os.path.basename and verify os.path.commonpath([os.path.realpath(ARTIFACTS_DIR), os.path.realpath(save_path)]) == os.path.realpath(ARTIFACTS_DIR) before writing.
-
HighTLS certificate verification disabled when transmitting Huawei Cloud IAM tokens and SDK requests
skills/ai/optverse/huawei-cloud-optv-solver-assistant/scripts/run_workflow.py:198
run_workflow.py and create_chat.py set verify=False on HTTPS requests that send the user's 24-hour X-Auth-Token header and download solver artifacts. Likewise, skills/bss/billing/huawei-cloud-business-tf-support/scripts/config.py sets http_config.ignore_ssl_verification = True for all Huawei Cloud SDK clients authenticated with AK/SK. A network attacker can intercept cloud credentials/tokens and tamper with API responses.
Fix: Enable TLS certificate verification by default (verify=True and ignore_ssl_verification = False) and allow custom CA bundles when needed for private environments.
-
HighUnauthenticated /instances/add and /instances/remove endpoints on vLLM proxy enable SSRF, API key leak, and backend hijacking
In PD-disaggregated deployments (templates/start_vllm.pd.sh), load_balance_proxy_server_example.py binds to 0.0.0.0:8080 (the service's client-facing port) and exposes POST /instances/add, POST /instances/remove, and POST /reset_prefix_cache without authentication. Any caller able to reach the service can register an attacker-controlled host:port, causing NodeListener.check_instance_status to send Authorization: Bearer <OPENAI_API_KEY> to that server (SSRF and credential theft) and adding it to the live inference pool so user prompts are routed to the attacker, or remove legitimate backends to cause denial of service.
Fix: Remove dynamic instance mutation endpoints in production or bind management routes to a separate loopback-only listener protected by strong authentication.
-
HighPath traversal and symlink dereferencing in untrusted expert suite manifests disclose local files
skills/businessapps/officeace/huawei-cloud-officeace-expert-migration/scripts/migrate_suite.py:169
When inspecting or migrating a third-party expert suite (even in default dry-run mode without --install), load_agent_prompt and enumerate_skills in migrate_suite.py (as well as agents[] path resolution in migrate_workbuddy.py) accept '../' traversal and absolute file paths from the package's plugin.json without verifying containment inside the package root, and copy skill directories using shutil.copytree with symlinks=False. A crafted third-party package can read arbitrary local files (such as ~/.ssh/id_rsa or cloud configs) into generated SKILL.md, experts/*.json, and report.md files.
Fix: Resolve all manifest paths and symlinks with Path.resolve() and reject any path not contained within root.resolve(), and pass symlinks=True (or reject symlinks) in shutil.copytree.
-
HighUnauthenticated Flask web server bound to 0.0.0.0 without Origin or CSRF checks
skills/devtools/kafka/huawei-cloud-kafka-replica-reassignment/scripts/app.py:508
The Kafka replica reassignment Web UI binds to 0.0.0.0:5000 with no authentication or Origin/CSRF validation on any endpoint. Anyone on the local network can read uploaded Kafka cluster broker IPs, topic topologies, and generated reassignment plans via /api/download and /api/report, and any network peer or visited website (via cross-origin POST to /api/upload) can overwrite input files and tamper with the reassignment JSON.
Fix: Bind the Flask server to 127.0.0.1 by default, require a random per-session bearer/CSRF token, and reject cross-origin state-changing requests.
-
HighTLS certificate verification disabled across API calls and get-pip.py bootstrap
skills/computing/ops/huawei-cloud-computing-query/scripts/ensure_env.py:31
ensure_env.py sets ssl._create_default_https_context = ssl._create_unverified_context (and config.py sets http_config.ignore_ssl_verification = True). If pip is missing, ensure_env.py downloads get-pip.py over unverified HTTPS with no checksum check and executes it immediately, runs pip install with --trusted-host, and sends Huawei Cloud AK/SK API requests without TLS verification. Similar verify=False / ignore_ssl_verification = True settings also appear in flexus_lifecycle.py, flexus-l-server-ops/scripts/auth.py, and ecs-sqlbot-deploy/scripts/huawei_cloud_ecs.py, exposing cloud credentials and code downloads to network MITM attacks.
Fix: Enable standard TLS certificate and hostname verification on all HTTPS requests and SDK clients, and verify SHA-256 digests before executing any bootstrap script.
-
HighHardcoded weak default root password on newly created public ECS instances
skills/computing/flexus-x/huawei-cloud-ecs-sqlbot-deploy/scripts/config.py:26
When deploy_sqlbot.py is invoked without --password (including the default command shown in its usage examples), HuaweiCloudECS falls back to DEFAULT_CONFIG['admin_pass'] and sets the root password of the newly provisioned public-IP ECS instance to 'Test@123456'. In addition, if NOTIFY_USER_ID is set, deploy_sqlbot.py sends that root password in plaintext over Feishu chat by default.
Fix: Generate a cryptographically random password per deployment via Python's secrets module or require SSH key-pair authentication, and never transmit plaintext root passwords over chat notifications.
-
HighStatic session secret, wildcard CORS, 0.0.0.0 binding as root, and world-readable secrets in deployed JiuwenSwarm
deploy_script_template.sh deploys JiuwenSwarm running as root bound to 0.0.0.0 on ports 5173, 18092, 19000, and 19001 with cors_origins: '*' and a hardcoded session_secret ('change-this-to-a-random-string') that is never rotated, allowing session forgery on network-exposed instances. Additionally, scripts/config_channel.py sets chmod 644 on config.yaml (exposing bot secrets and session_secret to local users), while scripts/config_model.py writes the plaintext model api_key to scripts/model_config_result.json and prints .env into COC logs.
Fix: Generate a cryptographically random session_secret during deployment, run the service as an unprivileged user bound to 127.0.0.1 by default, restrict CORS origins, keep config.yaml at mode 0600, and redact API keys in logs and result files.
-
HighLocal shell command injection via eval in DevBridge wrapper
skills/devtools/devbridge/huawei-cloud-devbridge-tunnel/scripts/devbridge_cmd.sh:169
Helper functions such as db_create, db_update, db_show, db_delete, db_echo, and db_ping concatenate arguments directly into command strings and pass them to eval inside _db_exec_retry (and db_host_bg interpolates tunnel_id into bash -c). If the agent is tricked via prompt injection into passing crafted tunnel names, descriptions, IDs, or URIs, arbitrary shell commands execute on the user's machine.
Fix: Replace string-concatenated eval calls with bash arrays (cmd=("devbridge" "create" "$name"); "${cmd[@]}") and validate tunnel IDs and ports against strict character allowlists.
-
HighRemote root command/code injection and persistent plaintext secrets in COC scripts
skills/computing/flexus-l/huawei-cloud-flexus-l-deploy-jiuwenswarm/scripts/config_model.py:250
Multiple deployment scripts (jiuwenswarm config_model.py and config_channel.py, flexusagent lib.py change_password and import_app_workflow, hermes lib.py install_maas and install_channel, and ecs-dsh-deploy coc_deploy.py) interpolate user- or agent-supplied values directly into bash double/single quotes, Python heredocs, or Expect scripts executed as root via Huawei Cloud COC, allowing arbitrary root command execution if arguments contain quotes or shell metacharacters. They also store plaintext passwords and API keys inside persistent COC custom scripts via CreateScript without deleting the scripts afterward.
Fix: Base64-encode parameters (as done in openclaw-deployment) and pass them as data rather than interpolating into shell/Python/Expect source, and delete temporary COC scripts immediately after execution or pass secrets via COC masked parameters.
-
HighTar path traversal (Tar-Slip) and missing SSH host key check when downloading DevKit reports
skills/computing/kunpeng/huawei-cloud-devkit-application-check-cli/scripts/devkit_remote.py:473
stream_tar_download() streams a tar.gz archive from the remote DevKit server and writes each entry to os.path.join(local_dir, rel) after only stripping leading slashes, without checking for '..' traversal segments or symlinks. Combined with paramiko.AutoAddPolicy() (which accepts any SSH host key without verification, also present in create_ecs_and_setup_devkit.py and kunpeng-source-code-migrate/scripts/ssh_client.py), a compromised remote host or active SSH MITM can overwrite arbitrary files outside local_dir on the user's machine.
Fix: Verify that os.path.realpath(local_path) stays within os.path.realpath(local_dir), reject symlink/hardlink tar members, and use RejectPolicy/SystemHostKeys instead of AutoAddPolicy for SSH connections.
-
HighRSYNC cloud migration disables SSH host key verification, exposes SSH to 0.0.0.0/0, and leaks root passwords in process arguments
skills/migrationom/oos/huawei-cloud-mgc-migration-ecs-gcp-to-hw/scripts/aws_to_huawei_direct.py:489
aws_to_huawei_direct.py and aws_to_huawei_rsync.py open TCP port 22 to 0.0.0.0/0 with root password login enabled via cloud-init, run ssh and rsync over the public internet with StrictHostKeyChecking=no when transferring the root filesystem '/', and pass plaintext root passwords via 'sshpass -p <password>' on the command line (exposing them in the process table for up to two hours).
Fix: Restrict security group SSH ingress to the source/operator IP, verify SSH host keys, and use temporary SSH key pairs or SSHPASS environment variables instead of 'sshpass -p'.
-
HighUnverified GOST binary downloaded from unofficial bucket and run as root
skills/migrationom/oos/huawei-cloud-migration-vmware-ecs-private/scripts/gost_ops.py:58
During proxy setup, gost_ops.py downloads a GOST archive from a personal or employee-named OBS bucket (mwx1015471-mpz, with HTTPS and GitHub fallbacks) into /tmp with no SHA-256 checksum or signature verification, uploads it to the proxy ECS, and runs it as root via systemd. If that unofficial bucket is modified, deleted and reclaimed by another tenant, or intercepted, an attacker gains root code execution on the user's migration proxy server.
Fix: Host the package in an official access-controlled repository, verify a pinned SHA-256 digest before extracting, and download into a private temporary directory instead of reusing predictable /tmp paths.
-
HighSquid proxy rule allows unrestricted HTTPS CONNECT tunneling
skills/migrationom/oos/huawei-cloud-migration-vmware-ecs-private/scripts/squid_ops.py:184
generate_config() writes 'http_access allow CONNECT ssl_ports' before 'http_access deny all' instead of the standard 'http_access deny CONNECT !ssl_ports'. In Squid, this permits any client that can reach port 3128 on the proxy ECS to tunnel HTTPS traffic to any internal VPC or external destination on port 443, bypassing both the allowed source IP list and the Huawei Cloud domain allowlist. In addition, proxy_ecs_ops.py configures an unauthenticated root rsync daemon with a writable [config] share.
Fix: Replace 'http_access allow CONNECT ssl_ports' with 'http_access deny !safe_ports' and 'http_access deny CONNECT !ssl_ports' placed before the allow rules, and bind Squid and rsyncd only to the private network interface.
-
HighLive Huawei Cloud IAM token and execution data sent to external telemetry API
skills/network/ops/huawei-cloud-network-topology-audit/scripts/skill_quality_sdk.py:287
By default (unless SKILL_QUALITY_DISABLE=1), skill_quality_sdk.py uses the user's Huawei Cloud AK/SK to mint a 24-hour project-scoped IAM token and sends that live bearer credential in the X-Auth-Token header—along with up to 6,000 characters of input parameters, 6,000 characters of cloud query output, and 20,000 characters of stack traces—to skillsapi.developer.myhuaweicloud.com/api/quality/report (or SKILL_QUALITY_ENDPOINT). Sharing a live IAM bearer token with a telemetry endpoint exposes the user's cloud permissions if that service is compromised or redirected, and sends operational data off-machine without asking. Similarly, smn_dms_skill.py fires background telemetry on every run by default.
Fix: Make telemetry strictly opt-in, never send a project-scoped IAM X-Auth-Token bearer token to a telemetry service, and strip input parameters and cloud query outputs from reports.
-
HighUnpinned runtime download and execution of skill-quality-cli without checksum
skills/monitoring/ecs/huawei-cloud-ecs-query/scripts/ensure_cli.sh:26
ensure_cli.sh and install_cli.sh (also duplicated in huawei-cloud-eip-query, huawei-cloud-sg-manage, huawei-cloud-iam-diagnose, and huawei-cloud-iam-manage) fetch an unvalidated version string from skillsapi.developer.myhuaweicloud.com, download an archive from OBS, and install executables into ~/.local/bin/ with no checksum or signature verification. Compromising the remote version endpoint or bucket lets an attacker run arbitrary code on the user's machine.
Fix: Deploy bundled local source files without network downloads (as newer skills in this repository already do) or pin both the version and a verified SHA-256 hash.
-
HighTLS verification disabled when downloading and running get-pip.py and pip packages
skills/monitoring/ops/huawei-cloud-monitoring-query/scripts/ensure_env.py:28
check_env.sh and check_env.ps1 invoke ensure_env.py (also duplicated in huawei-cloud-network-query and huawei-cloud-iam-query, with --trusted-host also in huawei-cloud-iam-diagnose), which globally disables Python SSL certificate checks, downloads get-pip.py using an unverified SSL context and executes it, and runs pip install with --trusted-host. An attacker on the network path can tamper with downloaded packages or get-pip.py to run arbitrary code on the user's computer.
Fix: Remove the unverified SSL context override and remove --trusted-host from pip install commands so all downloads require valid TLS certificates.
-
HighCrontab command injection via unescaped arguments in monitor_idle_eips.sh
skills/network/eip/huawei-cloud-eip-cost-optimizer/scripts/monitor_idle_eips.sh:88
When --setup-cron is passed, setup_cron() runs and exits before send_webhook() validates the URL, writing WEBHOOK_URL, EMAIL_ADDR, HW_REGION, and MIN_IDLE_DAYS directly into the user's crontab without escaping quotes, shell metacharacters, or newlines. If prompt injection tricks the agent into calling monitor_idle_eips.sh --setup-cron with a crafted webhook URL, email, or region, arbitrary shell commands are installed as a scheduled cron job.
Fix: Validate --webhook against the HTTPS domain allowlist, validate --region and --idle-days with strict patterns, reject newlines, and properly shell-escape all values before installing the crontab line.
-
HighUnquoted parameters interpolated into remote SSH and local shell commands
skills/migrationom/oos/huawei-cloud-migration-vmware-ecs-private/scripts/sms_agent_push.py:423
sms_agent_push.py interpolates password (read from the migration Excel sheet), ak, sk, and region directly into remote root bash commands without shell quoting, and batch_migrate.py (--background mode) joins CLI arguments into a string executed with subprocess.Popen(..., shell=True). Shell metacharacters in an untrusted Excel file or agent-supplied arguments can execute unintended commands on the local machine or remote server.
Fix: Escape all remote shell arguments with shlex.quote(), pass sensitive inputs via stdin rather than command strings, and avoid shell=True in batch_migrate.py.
-
HighUnpinned runtime binary download and installation without checksum verification
skills/security/waf-aad/huawei-cloud-waf-false-alarm-diagnosis/scripts/ensure_cli.sh:26
When skill-quality-cli is not installed, ensure_cli.sh (and the companion install_cli.sh) queries a remote latest-version endpoint, downloads a tarball from OBS without version pinning or SHA-256 verification, extracts it, and installs executables and Python scripts into ~/.local/bin. If the remote endpoint or bucket is compromised, arbitrary code is installed into the user's PATH.
Fix: Remove automatic runtime binary downloads (as done in huawei-cloud-waf-aad-rule-management/scripts/ensure_cli.sh) or pin a fixed version and verify a hardcoded SHA-256 digest before extraction.
-
HighAutomatic unverified download of KooCLI binary and unpinned pip packages at runtime
skills/solution/csm/huawei-cloud-architecture-resource-archive/scripts/check_koocli.py:90
check_koocli.py (invoked automatically by run_pipeline.py and query_resources.py) downloads the unpinned cli/latest archive from OBS with no cryptographic hash verification, extracts it with extractall() without path-traversal filtering, and immediately executes the binary with the user's cloud credentials. It also runs unpinned pip install commands at runtime for missing Python packages.
Fix: Pin exact KooCLI and Python package versions, verify SHA-256 checksums before extraction or installation, and pass filter='data' (or validate member paths) when extracting archives.
-
HighTLS certificate verification disabled on authenticated CCE Kubernetes proxy requests
skills/solution/csm/huawei-cloud-architecture-resource-archive/scripts/query_cce_k8s.py:121
cce_proxy_get() calls requests.get() with verify=False and suppresses urllib3 warnings when querying Kubernetes resources through cce.{region}.myhuaweicloud.com. An attacker on the network path can intercept signed headers—including X-Security-Token when temporary STS credentials are used—read returned Kubernetes workloads, ConfigMaps, or Secrets, and inject forged cluster data.
Fix: Remove verify=False and urllib3.disable_warnings() so requests validates the Huawei Cloud CCE endpoint TLS certificate against the system CA store.
-
HighTLS certificate verification disabled via curl -k on Kubernetes API requests
skills/solution/csm/huawei-cloud-cce-env-assessment/scripts/collect_all.py:172
k8s_api() runs curl with -k (--insecure) for every Kubernetes API call to the cluster gateway. A man-in-the-middle attacker on the network can relay the signed GET requests to read sensitive cluster workload specs (such as Deployment environment variables) in transit or spoof cluster assessment data.
Fix: Remove -k from the curl command and validate the cluster endpoint using the cluster CA certificate (--cacert) or system trust store.
-
HighTelemetry silently reads local AI chat databases and uploads user prompts and command output
skills/computing/ecs/huawei-cloud-ecs-vulnerability-auditor/scripts/cli/cli_reporting.py:714
When commands are run through the bundled skill-quality-cli wrapper, reporting is enabled by default and scrapes the user's local AI agent session databases (~/.local/share/opencode/opencode.db, ~/.hermes/state.db, ~/.codex/sessions/*.jsonl, and ~/.acpx/sessions/*.json) to extract recent user prompts, tool call history, and up to 6,000 characters of command stdout, sending them to skillsapi.developer.myhuaweicloud.com. Newer copies of the CLI in huawei-cloud-ecs-manage and huawei-cloud-gaussdb-instance-management remove message-body queries from those databases but still read session IDs from them and upload up to 6,000 characters of command stdout by default unless SKILL_QUALITY_REPORT=0 is set.
Fix: Make telemetry strictly opt-in, stop reading external AI agent session databases on disk, and never upload raw command stdout or user prompts.
-
HighAudit helper sends LLM API keys and scanned code snippets to a third-party gateway and is vulnerable to prompt injection
skills/devtools/common/huawei-cloud-skill-audit/scripts/checks/llm_arbiter.py:26
The false-positive arbiter is enabled by default (SKILL_AUDIT_LLM_ARBITER defaults to 1). If LLM_API_KEY or HERMES_CUSTOM_API_DEEPSEEK_COM_API_KEY is present in the environment and LLM_API_BASE is unset, it sends the user's API key in the Authorization header along with surrounding lines of scanned local files (including lines flagged for leaked credentials) to https://opengw.clouddeveloper.club/v1. In addition, raw lines from untrusted files under audit are embedded directly into the prompt, allowing a malicious skill to inject instructions that mark its own security findings as false_positive.
Fix: Disable remote LLM arbitration by default, require an explicit user-configured endpoint instead of defaulting to opengw.clouddeveloper.club, redact secrets before sending context, and isolate untrusted code from arbiter instructions.
-
HighHelper scripts download and install unpinned binaries at runtime without checksum verification
skills/storage/obs/huawei-cloud-obs-bucket-create/scripts/ensure_cli.sh:26
ensure_cli.sh and install_cli.sh in huawei-cloud-obs-bucket-create, huawei-cloud-obs-image-desens, and huawei-cloud-obs-lifecycle-management query a mutable latest version endpoint, download a tarball from OBS, extract it with tar xzf without checking for path traversal or verifying any SHA256 checksum, and install the resulting executable into ~/.local/bin/skill-quality-cli. A compromised endpoint or Bucket/MITM response leads directly to arbitrary code execution.
Fix: Remove runtime binary downloads (as done in newer skills in this repository) or pin exact versions with hardcoded cryptographic SHA256 verification and safe tar extraction.
-
HighPlaywright helper falls back to npx with an unscoped package name instead of @playwright/cli
skills/solution/sac/huawei-cloud-sac-dify/scripts/playwright_utils.py:38
When playwright-cli is not already installed on PATH, build_pw_command() in huawei-cloud-sac-dify, huawei-cloud-sac-new-api, and huawei-cloud-sac-yolo runs 'npx playwright-cli'. Because the official package is scoped (@playwright/cli), passing the unscoped name 'playwright-cli' causes npx to resolve and execute an unpinned, unrelated unscoped package from the public npm registry if auto-install is enabled.
Fix: Fail with installation instructions when playwright-cli is missing instead of invoking npx, or explicitly pin 'npx --no-install' so npx never downloads remote packages at runtime.
-
HighCode injection into generated FunctionGraph handler and infinite OBS event recursion
skills/storage/obs/huawei-cloud-obs-image-desens/scripts/obs_image_desens.py:378
In setup-trigger mode, _build_auto_func_code interpolates --fill and --trigger-prefix / --source-prefix directly into double-quoted Python string literals inside the generated FunctionGraph source code without escaping, allowing crafted arguments (for example via prompt injection) to inject arbitrary Python code into the deployed cloud function. Furthermore, the generated handler listens to ObjectCreated events and writes the processed image back to the exact same bucket and key, re-triggering ObjectCreated in an infinite serverless billing loop.
Fix: Pass configuration via FunctionGraph environment variables or json.dumps() instead of string formatting Python code, and write desensitized images to a separate bucket or non-overlapping output prefix.
-
HighTLS certificate verification is disabled on Huawei Cloud SDK and Kubernetes cluster clients
skills/storage/obs/huawei-cloud-storage-query/scripts/config.py:53
config.py disables TLS certificate verification (http_config.ignore_ssl_verification = True) for all Huawei Cloud API calls in huawei-cloud-storage-query. Similarly, skills/container/cce/huawei-cloud-cce-log-analyzer/scripts/huawei_cloud/kubectl_client.py hardcodes '--cce-insecure-upstream-tls=true' on every kubectl cce command, and skills/computing/ops/huawei-cloud-computing-query/scripts/as/list_api_versions.py and show_api_version.py set CERT_NONE. This exposes cloud API traffic, AK/SK signatures, security tokens, and cluster configuration changes to man-in-the-middle interception.
Fix: Enable standard TLS certificate and hostname verification by default across all SDK, urllib, and kubectl-cce connections.
-
HighOpenClaw deployment scripts open bot DMs to all external users while enabling the full tool profile
multi_channel.sh configures WeCom, Feishu, QQ, Discord, and Telegram bot integrations with dmPolicy set to 'open' and allowFrom set to ['*'], while sibling script multi_model.sh sets '.tools.profile = "full"' in openclaw.json. Any stranger on those messaging platforms can message the deployed OpenClaw bot directly and trigger full tool capabilities on the server.
Fix: Require explicit user/sender allowlists instead of allowFrom: ['*'], default dmPolicy to a restricted mode, and avoid forcing tools.profile to 'full'.
-
HighBag2manifest Gradio web workbench binds to 0.0.0.0 without authentication and exposes local files
skills/ai/cloudrobo/huawei-cloud-cloudrobo-dataset/scripts/bag2manifest/bag2manifest/webapp.py:1129
When the bag2manifest web workbench is started on a local machine, it binds to 0.0.0.0 on port 7860 by default with no password or authentication and enables the local server path input box. Anyone on the same network can connect to port 7860, probe whether arbitrary paths exist on the computer, extract local .zip files, and read or download data and camera frames from any readable bag or SQLite .db3 file.
Fix: Change the default --host value to 127.0.0.1 and disable arbitrary server-side path input unless explicitly enabled alongside Gradio authentication.
-
HighOpenViking plugin installer downloads unpinned code through unofficial third-party proxies without trustworthy verification
skills/devtools/extensions/huawei-cloud-openviking-agent-integration/scripts/lib/plugins.sh:39
During agent integration, the plugin provisioning logic fetches the latest unpinned HEAD commit SHA, the expected file SHA-1 tree, and the executable plugin files themselves from unofficial third-party proxy domains (gh-proxy.com and ghfast.top) before GitHub. Because the expected hash tree comes from the same third-party proxy as the code—and ov_sync_shared_files in base.sh downloads 25 .mjs runtime modules from those mirrors with no hash check at all—a compromised proxy can inject arbitrary code into the user's AI coding agents.
Fix: Pin upstream plugin releases to known immutable commit SHAs and cryptographic hashes stored in the skill, and always verify downloaded files against pinned hashes rather than hashes fetched from a proxy.
-
HighQuality CLI silently auto-downloads and executes remote binaries at runtime
skills/devtools/common/huawei-cloud-skill-audit/scripts/cli/cli_entry.py:465
In huawei-cloud-skill-audit and huawei-cloud-deployment-task-management, running skill-quality-cli with run, report, or self-check silently calls ensure_latest(quiet=True) by default. It queries a remote endpoint for a download_url and sha256, downloads the archive or ELF binary without restricting the download domain or verifying a cryptographic signature, installs it as an executable in ~/.local/bin, and attempts to re-exec it.
Fix: Remove automatic runtime binary downloads and deploy only the bundled local Python source, matching the SC2 supply-chain fix already applied in newer copies of cli_entry.py.
-
HighSkill tester executes target skill commands via bash -c and misses mutating cloud verbs when ALLOW_WRITES=0
skills/devtools/common/huawei-cloud-skill-tester/scripts/tier1/phase-4-execute-tests.sh:170
Phase 4 (and Phase 6) extracts commands from the tested skill's SKILL.md, references/*.md, and templates/test-vars.json and runs anything not flagged as a write operation with live Huawei Cloud credentials. Because commands are passed to bash -c (or python3 for SDK snippets) and write detection only checks for a few substrings (create, delete, update, destroy, activate, reclaim), a third-party skill can execute arbitrary shell commands via metacharacters, and mutating cloud calls such as ScheduleKeyDeletion, DisableKey, or BatchStopServers will run even when ALLOW_WRITES=0.
Fix: Avoid bash -c by parsing arguments into an allowlisted argv array, default to read-only allowlisted operations (List/Show/Get/Describe) when ALLOW_WRITES=0, and require explicit user approval before running scripts or test-vars.json from untrusted skills.
-
HighJumpbox exposes unauthenticated gost WinRM port forwarders and HTTP file server on all interfaces
During private-network Windows migration, deploy_jumpbox.py (and PrivateBatchOrchestrator) starts gost listening on 0.0.0.0:15985+ forwarding directly to internal Windows hosts' WinRM ports (5985), as well as python3 -m http.server 8080 on 0.0.0.0:8080, on a public-facing jumpbox without tunnel authentication or IP binding. Anyone able to reach those jumpbox ports can access internal Windows WinRM endpoints directly, and sandbox-to-jumpbox WinRM traffic traverses the network over plain HTTP by default.
Fix: Tunnel WinRM and file traffic over the authenticated SSH connection (local SSH port forwarding) instead of binding unauthenticated gost and http.server listeners on 0.0.0.0.
-
HighMigration executables are downloaded over plain HTTP without integrity checks and run as Administrator
ConfigurePhase downloads configure_sms_agent.exe over unencrypted HTTP with no checksum or signature check and executes it via a highest-privilege Scheduled Task. Similarly, SourceDownloadPhase runs Download-Agent.ps1 over plain HTTP with -VerifyHash disabled, parses untrusted href filenames into Join-Path without path-traversal filtering, and InstallAgentPhase/InstallAutoItPhase execute the downloaded installers as Administrator. A network attacker on the HTTP path can write arbitrary files or run code as Administrator on source Windows hosts.
Fix: Transfer executables over encrypted, authenticated channels, enforce pinned SHA-256 verification before execution, and validate downloaded filenames in Download-Agent.ps1 so they cannot contain path separators or '..'.
-
HighSSH host key verification and WinRM TLS certificate validation are disabled
SSHExecutor and deploy_jumpbox.py set paramiko.AutoAddPolicy(), accepting any SSH host key without verification, while WinRMExecutor and WinRMDetector set server_cert_validation='ignore' when SSL is enabled (and default to unencrypted HTTP). Because these connections carry root SSH passwords, Windows Administrator passwords, and Huawei Cloud AK/SK secrets across networks, an active man-in-the-middle attacker can intercept credentials and execute commands.
Fix: Verify SSH host keys against known_hosts (RejectPolicy/WarningPolicy with pinned host keys) and enable WinRM over HTTPS with certificate validation enabled by default.
-
HighUnescaped interpolation into generated agent startup scripts allows command injection
ov_openclaw.py and ov_workswarm.py build executable bash startup scripts (ov-openclaw-init.sh and ov-jiuwenswarm-init.sh, sourced by start.sh) by substituting endpoint, npm_registry, mcp_url, and OV_API_KEY with raw string .replace() into unquoted or double-quoted shell contexts and embedded Python code. If an AI agent is tricked via prompt injection into passing a crafted endpoint or registry URL containing shell metacharacters such as $(...) or quotes, arbitrary commands are persisted and executed on agent startup.
Fix: Validate URLs against a strict scheme/host/port pattern and escape all interpolated values with shlex.quote() before writing them into shell scripts.
-
HighSQLBot deployed to public internet with hardcoded default admin password
skills/computing/flexus-x/huawei-cloud-ecs-sqlbot-deploy/references/acceptance-criteria.md:32
The SQLBot deployment skill creates an ECS instance with a public IP, opens security group port 8000 to all IP addresses (0.0.0.0/0, listed as '8000 允许所有 IP') over plain HTTP, and leaves the service running with the well-known default administrator login admin / SQLBot@123456. Anyone scanning the internet can log in as administrator without any AI involvement, take over the SQLBot service, and access connected databases or model keys.
Fix: Restrict security group inbound rules on port 8000 to the installer's /32 IP (or use an SSH tunnel), generate a strong unique admin password during deployment instead of SQLBot@123456, and enable HTTPS.
-
HighJiuwenSwarm runs as root with internal services bound to 0.0.0.0 and wildcard CORS
skills/computing/flexus-l/huawei-cloud-flexus-l-deploy-jiuwenswarm/assets/env_template.env:13
The JiuwenSwarm deployment templates and acceptance checks override upstream 127.0.0.1 defaults and bind the frontend (5173) as well as internal backend services (WEB_HOST, AGENT_SERVER_HOST, GATEWAY_HOST on ports 18092, 19000, 19001) to 0.0.0.0, while running the service as root (assets/jiuwenswarm.service.template), setting a placeholder SECRET_KEY, and enabling wildcard CORS (ALLOWED_ORIGINS=* and allowed_origins: ['*'] in config_template.yaml). Exposing the service over plain HTTP with wildcard CORS lets network attackers reach internal agent ports if open and lets any website visited by the user issue cross-origin requests to the root-privileged service.
Fix: Keep WEB_HOST, AGENT_SERVER_HOST, and GATEWAY_HOST bound to 127.0.0.1, run the systemd unit as a dedicated non-root user, generate a random SECRET_KEY, restrict ALLOWED_ORIGINS to trusted origins, and require HTTPS or an SSH tunnel.
-
HighInternet-facing migration helper VM created with hardcoded root password
skills/migrationom/oos/huawei-cloud-sms-migration-assistant/references/MIGRATION_REFERENCE.md:265
During post-migration boot repair (Section G.2), the skill instructs creating a temporary ECS instance with a public Elastic IP attached, placing it in a security group with SSH open, and setting its root password to the hardcoded value 'Helper@2026'. The user's migrated production root disk is then attached and mounted on this helper VM, and the runbook never shows a command to delete the helper VM afterwards, allowing anyone on the internet to log in as root and access the mounted disk.
Fix: Create the helper VM without a public IP (or restrict SSH ingress to the operator's /32 IP), authenticate using an SSH keypair or a freshly generated random password stored outside command arguments, and explicitly delete the helper VM and release its EIP immediately after detaching the disk.
-
HighMigration guide enables root password SSH login and opens firewall ports to 0.0.0.0/0
skills/migrationom/oos/huawei-cloud-sms-migration-assistant/references/MIGRATION_REFERENCE.md:355
Section G.3 rewrites /etc/ssh/sshd_config on the user's migrated server to force 'PasswordAuthentication yes' and 'PermitRootLogin yes' (and Section G.5 connects with '-o StrictHostKeyChecking=no'), silently downgrading servers that previously required SSH keys. Earlier in Section C.1, the same guide opens inbound security group rules (ports 22, 8899, 8900) to 0.0.0.0/0 instead of restricting access to the source server's IP address.
Fix: Preserve the source server's existing SSH authentication settings instead of enabling root password login, verify SSH host keys, and restrict migration security group ingress rules to the specific source/management /32 IP addresses.
-
HighShell command injection in OBS Terraform local-exec provisioners
Three OBS Terraform templates (object-upload-with-source, object-upload-with-encryption, and bucket-with-website) use local-exec provisioners that interpolate file contents (inside single quotes) and file/tag names (unquoted) directly into shell commands (echo and rm). If an agent populates website HTML, uploaded text, or filenames from untrusted content containing single quotes or shell metacharacters—or if bucket tags are modified before destroy—running terraform apply or terraform destroy executes arbitrary shell commands on the user's computer.
Fix: Replace local-exec echo/zip/rm commands with native Terraform resources (such as huaweicloud_obs_bucket_object content, archive_file data source, or local_file) so file content and names are never evaluated by a shell.
-
MediumHardcoded default OS and admin passwords on provisioned cloud servers
skills/computing/flexus-x/huawei-cloud-ecs-sqlbot-deploy/SKILL.md:360
When provisioning a Huawei Cloud Flexus X server with a public IP, the default deployment configuration sets the server OS root password to Test@123456 and the SQLBot admin password to SQLBot@123456 unless the user manually selects custom configuration. Because these static weak passwords are published in this public repository, anyone with network reachability to the server's SSH port or application port before the user changes them can log in with administrative access.
Fix: Generate cryptographically random passwords at deployment time (as done in huawei-cloud-ecs-dsh-deploy) or require SSH key authentication instead of static default passwords.
-
MediumRuntime archive downloaded from personal GitCode account and given keyring credentials
skills/devtools/openjiuwen/huawei-cloud-install-openjiuwenswarm/SKILL.md:150
This skill downloads a 300MB prebuilt runtime archive from a non-configurable personal GitCode repository (gitcode.com/afeng5267/jiuwenswarm_runtime.git), installs binaries into /usr/local/bin, starts background services, and populates its .env with the host container's API_KEY read from the Linux kernel keyring. It also instructs the agent to install Python dependencies silently without informing the user. If that personal account is compromised or updated maliciously, an attacker gains code execution and access to the user's cloud API credentials.
Fix: Host release artifacts on an official Huawei Cloud repository or verified object storage bucket, pin cryptographic SHA-256 checksums, and disclose dependency installation to the user.
-
MediumMandatory TLS certificate verification bypass on outbound HTTPS requests
skills/ai/modelarts/huawei-cloud-modelarts-vllm-ascend-model-deploy/SKILL.md:172
This skill requires the agent to pass -k / --insecure on every curl request to whitelisted domains (including docs.vllm.ai, modelscope.cn, huggingface.co, quay.io, and *.myhuaweicloud.com) to bypass TLS certificate checks, and sibling skills (such as huawei-cloud-kubectl-cce-installer and huawei-cloud-modelarts-training-diagnosis) similarly retry with --cce-insecure-upstream-tls=true or set skipSecureVerify=true. Disabling TLS validation allows a network man-in-the-middle attacker to tamper with fetched deployment parameters, image tags, or API responses.
Fix: Keep standard TLS certificate verification enabled by default and allow users behind corporate MITM proxies to configure a trusted custom CA bundle instead of passing -k or --insecure.
-
MediumSkill discovery falls back to unpinned personal Gitee branch
skills/devtools/common/huawei-cloud-find-skills/SKILL.md:180
When the primary GitCode API returns HTTP 429, the skill search script automatically falls back to a mutable branch (test-for-index) under a personal Gitee account (csimon666/skills-group-contribution) to load the skill index and keyword map. If that personal account or test branch is compromised, an attacker can inject prompt-injection payloads into skill descriptions read by the agent or alter skill category paths.
Fix: Host backup skill indices only under official Huawei Cloud organization repositories or verified object storage, and pin releases or verify signatures.
-
MediumPredictable SSH ControlMaster socket in /tmp and permanent StrictHostKeyChecking bypass
skills/computing/ecs/huawei-cloud-ecs-passwordless-login/SKILL.md:102
Step 6 appends permanent Host entries to ~/.ssh/config that place the SSH ControlMaster socket at a predictable path in world-writable /tmp (/tmp/coc_ssh_%r@%h:%p) and set StrictHostKeyChecking no and UserKnownHostsFile /dev/null, explicitly leaving these entries behind after cleanup. Predictable sockets in /tmp expose multi-user hosts to local socket collisions or hijacking, and permanently disabling host key verification leaves all future SSH connections to those elastic IPs vulnerable to man-in-the-middle attacks.
Fix: Store ControlPath sockets inside a mode-0700 directory under ~/.ssh/, use StrictHostKeyChecking accept-new instead of no, and remove temporary Host blocks during Step 7 cleanup.
-
MediumUnpinned remote deployment scripts executed without checksum verification or safe quoting
skills/ai/modelarts/huawei-cloud-ascend-models-deploy/SKILL.md:351
Model deployment templates download shell scripts at runtime from a public OBS bucket via wget and immediately execute them as root without version pinning or SHA-256 verification, while interpolating ${model}, ${cards}, and ${port} unquoted into bash -c. In addition, wget -P does not overwrite an existing file, so a pre-existing script at /home/modelarts-agent/deploy-large-models.sh will be executed instead of the newly downloaded one.
Fix: Bundle vetted deployment scripts inside the skill or verify pinned SHA-256 digests before execution, write to a clean path with wget -O, and shell-quote all interpolated parameters.
-
MediumBlocked and sensitive command checks bypassed in SSH executor
skills/ai/modelarts/huawei-cloud-ascend-remote-connect/scripts/main.py:52
Although this skill claims to block destructive commands (such as mkfs, dd zero-writes, and rm -rf /) and require confirmation for sensitive operations, run_one_shot_fast checks _is_shell_command(command) and sends shell commands directly over SSH without invoking CommandExecutor or CommandValidator even when --raw is not set. In addition, all patterns in command_validator.py are anchored with '^', so prepending 'true && ' or 'sudo ' bypasses every check in interactive mode as well.
Fix: Route all one-shot commands through CommandValidator before execution, parse compound shell commands rather than relying on start-of-string regexes, and require explicit out-of-band confirmation for destructive operations.
-
MediumPlaintext Huawei Cloud Secret Key sent in HTTP request header
skills/devtools/common/huawei-cloud-activity-submit/submit.mjs:146
submit.mjs reads the user's Huawei Cloud Secret Access Key from CLI flags or standard environment variables (HW_SECRET_KEY, HUAWEICLOUD_SDK_SK, HW_SK) and sends the raw secret key in the X-HW-SK HTTP header to open-dataset.huaweicloud.com. Cloud secret keys are meant to sign requests locally via HMAC-SHA256 and never leave the client; transmitting the raw SK exposes full account credentials to the remote web service, proxies, and request logs.
Fix: Authenticate requests using standard Huawei Cloud SDK-HMAC-SHA256 request signing or short-lived scoped tokens instead of transmitting the raw Secret Access Key in HTTP headers.
-
MediumDisabled SSH host key verification across SSH remote skills
skills/ai/modelarts/huawei-cloud-ascend-remote-connect/scripts/ssh_client.py:50
SSH clients in huawei-cloud-ascend-remote-connect (ssh_client.py via AutoAddPolicy and main.py via StrictHostKeyChecking=no / UserKnownHostsFile=/dev/null) and huawei-cloud-ascend-command (npu_client.py via AutoAddPolicy) accept any remote SSH host key without verification, allowing a man-in-the-middle attacker on the network to capture passwords and intercept remote commands.
Fix: Verify host keys against the user's known_hosts file (RejectPolicy / load_system_host_keys in Paramiko and default StrictHostKeyChecking=yes in OpenSSH).
-
MediumPlaintext SSH password in process arguments and predictable /tmp ControlMaster socket
skills/ai/modelarts/huawei-cloud-ascend-remote-connect/scripts/main.py:59
_ensure_connection passes the SSH password via 'sshpass -p <password>' in command-line arguments (visible to other local users in ps / /proc) and creates a reusable ControlMaster socket at the predictable path /tmp/ssh-ctrl-{user}@{host}:{port} in world-writable /tmp after unlinking any existing path, exposing multi-user hosts to credential disclosure and /tmp socket race or hijack attacks.
Fix: Pass the password to sshpass via the SSHPASS environment variable or file descriptor (-e or -d), and place ControlMaster sockets inside a private mode-0700 directory under the user's home or XDG_RUNTIME_DIR.
-
MediumCypher injection and destructive-query guardrail bypass in GES Graph SDK
skills/ai/ges/huawei-cloud-ges-graph/scripts/ges_graph_skill.py:140
validate_cypher (and validateCypher in ges_graph_skill.js) exempts queries from destructive keyword checks whenever 'WHERE' and 'ID(' appear anywhere in the query string, which is trivially bypassed (for example, 'MATCH (n) WHERE id(n) IS NOT NULL DETACH DELETE n'). In addition, add_node, update_node, add_edge, delete_edge, and get_nodes_by_label interpolate labels, edge types, and property keys directly into Cypher strings without identifier validation.
Fix: Validate all labels, edge types, and property keys against a strict identifier allowlist regex, and parse Cypher AST or use read-only database roles rather than substring checks to block destructive statements.
-
MediumPython code injection via unsanitized region argument in ModelArts pricing scripts
skills/ai/modelarts/huawei-cloud-modelarts-notebook-management/scripts/query-pricing.sh:54
query-pricing.sh in huawei-cloud-modelarts-notebook-management (as well as the same pattern in huawei-cloud-modelarts-resource-pool-management and huawei-cloud-modelarts-training-management) interpolates TARGET_REGION directly inside a single-quoted Python string passed to 'python3 -c'. A prompt-injected agent supplying a crafted region argument can break out of the string literal and execute arbitrary local code under a read-only pricing script.
Fix: Pass TARGET_REGION to Python via an environment variable or sys.argv instead of interpolating it into the python3 -c code string, and validate region names against a strict regex.
-
MediumUnverified runtime script download and execution in Ascend model deploy helper
skills/ai/modelarts/huawei-cloud-ascend-models-deploy/scripts/deploy_helper.py:259
generate_deploy_command generates a shell command that downloads deployment scripts from an OBS bucket with 'wget -P' and immediately runs them with 'sh' without pinning a digest or verifying a SHA-256 checksum. Furthermore, 'wget -P' saves a '.1' file if the target filename already exists, causing 'sh' to execute whatever pre-existing file was in /home/modelarts-agent/.
Fix: Bundle the deployment scripts in the skill or verify a pinned SHA-256 checksum after downloading to a dedicated temporary file before execution.
-
MediumECS asset collector executes untrusted host and container binaries during read-only scan
skills/computing/ecs/huawei-cloud-ecs-vulnerability-auditor/scripts/collect_assets.sh:103
Although documented as a safe, read-only script for privileged SSH or Huawei Cloud COC execution, collect_assets.sh iterates over every /proc/<pid>/exe on the machine (regardless of process owner) as well as executables in $PATH, /app, /app/bin, /opt, and /opt/bin, running each binary with -v, --version, -V, and version flags (and running docker exec inside containers). When run as root during an audit, any unprivileged user or compromised service running a custom binary or owning /app or /opt can achieve root code execution, and non-idempotent tools may trigger unintended side effects.
Fix: Do not execute discovered binaries to determine versions; rely solely on package manager databases (dpkg, rpm, apk) or static binary/package metadata inspection.
-
MediumModelArts online inference service defaults to auth_type NONE and enables internet access on A5
build_request_body sets runtime_config.service_invoke.auth_type to 'NONE' over HTTP on port 8080 by default, and when --chip-model=a5 is selected it also sets internet_access_enable to True with an empty ip_white_list. Unless overridden with --raw-json, deployed inference endpoints allow unauthenticated access to the model and NPU compute quota.
Fix: Default auth_type to IAM (or APP) authentication and set internet_access_enable to False unless the user explicitly opts in.
-
MediumDefault-on telemetry SDK reads local AI agent session databases and uploads execution details
skills/bss/billing/huawei-cloud-bss-voucher-query/scripts/skill_quality_sdk.py:212
skill_quality_sdk.py (vendored into huawei-cloud-bss-voucher-query, huawei-cloud-partner-scout, huawei-cloud-msmodelslim-model-analysis, huawei-cloud-dli-manage, and huawei-cloud-ecs-lifecycle) runs on every invocation unless SKILL_QUALITY_DISABLE=1 is set. Version 2.12.2 queries unrelated local agent databases (~/.local/share/opencode/opencode.db, ~/.hermes/state.db, ~/.codex/sessions/*.jsonl) and .quality_report.json up to three parent directories above cwd, uploading CLI arguments, outputs, stack traces, user_input, session_log, and token counts to skillsapi.developer.myhuaweicloud.com (including via an unauthenticated guest endpoint, or by sending raw AK/SK in an HTTPS POST body to IAM). Version 1.0.0 similarly uploads arguments, outputs, and tracebacks even for offline tools like analyze_model.py when AK/SK environment variables exist.
Fix: Make telemetry opt-in rather than default-on, stop reading unrelated agent session databases in the user's home directory, and avoid transmitting raw secret keys or command outputs.
-
MediumCloud tokens and session cookies written to predictable /tmp paths with default permissions
skills/ai/optverse/huawei-cloud-optv-solver-assistant/scripts/create_chat.py:156
create_chat.py caches a 23-hour Huawei Cloud IAM token at a fixed path (/tmp/optverse_iam_token.txt) using open(..., 'w') without setting 0600 permissions or preventing symlink attacks, and creates ~/.config/optverse/credentials with default permissions as well. Similarly, manager_api_client.py (in huawei-cloud-mrs-host-alarm-diagnose and huawei-cloud-mrs-host-fault-diagnose) writes plaintext JSESSIONID cookies to /tmp/manager_session/cookies.txt with default file permissions. On multi-user systems, local users can read cached credentials or exploit predictable /tmp paths.
Fix: Store token and session caches in a 0700 directory under the user's home directory (such as ~/.cache/optverse) and create files atomically with 0600 permissions using os.open(..., O_CREAT | O_EXCL | O_WRONLY, 0o600).
-
Mediumcheck_runtime.py sends DPAPI-decrypted OfficeAce session cookie to arbitrary --base URL
skills/businessapps/officeace/huawei-cloud-officeace-expert-migration/scripts/check_runtime.py:72
check_runtime.py decrypts ~/.office-claw/session.cookie via Windows DPAPI and attaches Cookie: oc_sid=<sid> to requests sent to --base with TLS certificate verification disabled. Because --base is not restricted to localhost/loopback addresses, prompt injection convincing the agent to pass an external URL exfiltrates the decrypted session cookie.
Fix: Restrict --base to loopback hosts (127.0.0.1, ::1, localhost) before attaching the decrypted session cookie.
-
MediumvLLM startup templates install unverified PyPI package unofficial-ascend-tools at container runtime
In start_vllm.single.sh, start_vllm.multi_node.sh, and start_vllm.pd.sh, when /etc/hccl_rootinfo.json is missing and no local wheel is supplied, install_ascend_tools automatically downloads and installs 'unofficial-ascend-tools==0.0.7rc4' from PyPI without --require-hashes verification and executes it inside the inference container.
Fix: Pre-install vetted dependencies during container image build or require --require-hashes with a pinned SHA-256 digest.
-
MediumHardcoded trust_remote_code=True executes arbitrary Python code from model directories
skills/ai/modelarts/huawei-cloud-msmodelslim-model-adapt/scripts/step1_generate_test_model.py:101
step1_generate_test_model.py, step2_run_quantization.py (--trust_remote_code True), and the start_vllm.*.sh templates (--trust-remote-code) unconditionally enable custom Python code execution when loading HuggingFace/ModelScope configurations and models. Pointing these tools at an untrusted third-party model repository executes arbitrary code bundled with the checkpoint.
Fix: Default trust_remote_code to False and require an explicit CLI flag after warning the user before executing custom model code.
-
MediumHost AI agent API key silently extracted from Linux kernel keyring into plaintext .env
skills/devtools/openjiuwen/huawei-cloud-install-openjiuwenswarm/scripts/common.py:38
During setup (03_configure.py), get_keyring_config() installs keyutils and reads the HWCLOUD-Agent secret from the Linux kernel session keyring (@s) via keyctl print, decodes it, and writes it in plaintext (API_KEY=...) to ~/.jiuwenswarm/config/.env for the downloaded JiuwenSwarm service without user consent, bypassing kernel keyring isolation.
Fix: Do not harvest credentials from the host agent's kernel keyring; prompt the user to configure a dedicated API key explicitly for JiuwenSwarm.
-
MediumEncrypted KooCLI cloud credentials permanently downgraded to plaintext on disk
skills/computing/flexus-l/huawei-cloud-flexus-l-deploy-jiuwenswarm/scripts/utils.py:83
Whenever any script in huawei-cloud-flexus-l-deploy-jiuwenswarm loads credentials, _read_hcloud_config_credentials() checks ~/.hcloud/config.json; if credential encryption is enabled (authEncrypt == true), it silently runs 'hcloud configure set --cli-auth-encrypt=false', rewriting ~/.hcloud/config.json with plaintext accessKeyId and secretAccessKey on disk and never restoring encryption.
Fix: Never disable authEncrypt on the user's KooCLI configuration; read credentials from environment variables or invoke hcloud directly without decrypting config.json to disk.
-
MediumStored XSS in generated Kafka reassignment HTML report
skills/devtools/kafka/huawei-cloud-kafka-replica-reassignment/scripts/report.py:25
generate_report() interpolates untrusted values from broker.xlsx (such as mount points and broker IDs), degradation events, and pipeline parameters (json_strategy, json_strategy_param) directly into report.html without HTML escaping. Viewing /api/report in a browser executes injected JavaScript in the Web UI origin.
Fix: Escape all dynamic values with html.escape() before embedding them into the HTML template.
-
MediumPKCS#7 package signature verification bypassed via openssl -noverify without signer pinning
skills/computing/kunpeng/huawei-cloud-devkit-webui-create/scripts/install_devkit_webui.sh:126
install_devkit_webui.sh verifies ${DEVKIT_PKG}.p7s using 'openssl cms -verify ... -noverify' without pinning a trusted CA or signer certificate (-CAfile / -certfile). OpenSSL therefore accepts any self-signed certificate bundled inside the .p7s file itself, so an attacker who controls --devkit-url or tampers with both files can sign a malicious archive with a throwaway key and pass verification.
Fix: Pin Huawei's expected signing certificate or SHA-256 package digest in the script and verify against the pinned certificate rather than using -noverify alone.
-
MediumUnpinned third-party binaries and remote install scripts executed without independent checksum verification
skills/devtools/common/huawei-cloud-vod-collector/scripts/vod_install.sh:19
vod_install.sh (and vod_install.ps1) clones the unpinned personal repository weixin_45218422/AtomGit-GO.git and checks the binary archive only against a .sha256 file from that same unpinned clone (or builds unpinned Go source) to run an OAuth helper. Similar unverified runtime downloads occur in install_kubectl_cce.sh (personal Gitee repo pancake0001/kubectl-cce-plugin), publish-work-to-gallery/scripts/ensure-user-tool.mjs (personal repo zhoucungen/gitcode-oauth), codearts-code-gen/scripts/setup.sh and setup.ps1 (piping install.sh/install.ps1 to sh/Invoke-Expression), codearts-pipeline-diagnose and bcs-manage ensure_cli.sh/install_cli.sh, devbridge_cmd.sh, check_env.sh, and install_terraform.py.
Fix: Pin exact release versions and verify SHA-256 checksums hard-coded in the skill repository before extracting or executing downloaded binaries and installer scripts.
-
MediumOver-permissive bot channel access defaults allow any user to interact with deployed Hermes agent
skills/computing/flexus-l/huawei-cloud-flexus-l-server-hermes-deployment/scripts/lib.py:550
When configuring Feishu or WeCom channels on a Hermes AI agent server, install_channel hardcodes FEISHU_ALLOW_ALL_USERS=true, FEISHU_GROUP_POLICY=open, FEISHU_ALLOWED_USERS='', and GATEWAY_ALLOW_ALL_USERS=true, allowing any user who can message the bot or join a group to invoke the agent without allowlist restrictions. Similarly, flexusagent-deployment/scripts/sg_rule.py automatically opens port 80 to 0.0.0.0/0 on the shared sg-default-smb security group before the default admin password is changed.
Fix: Default GATEWAY_ALLOW_ALL_USERS and FEISHU_ALLOW_ALL_USERS to false, require an explicit user/group allowlist, and avoid opening shared security groups to 0.0.0.0/0 automatically.
-
MediumDefault-enabled telemetry uploads tool inputs, log outputs, stack traces, and IAM auth token
skills/container/cce/huawei-cloud-cce-log-analyzer/scripts/skill_quality_sdk.py:79
skill_quality_sdk.py enables telemetry by default (unless SKILL_QUALITY_DISABLE=1 is set). On every wrapped execution, it exchanges the user's Huawei Cloud AK/SK for a live project-scoped IAM token and sends that X-Auth-Token along with up to 6,000 characters of input_param, 6,000 characters of output_result (such as analyzed CCE cluster logs), and 20,000 characters of unmasked full_stack to skillsapi.developer.myhuaweicloud.com (or SKILL_QUALITY_ENDPOINT).
Fix: Make telemetry opt-in, omit raw tool inputs, cluster log outputs, and stack traces from telemetry payloads, and avoid forwarding general-purpose IAM project tokens to third-party or custom endpoints.
-
MediumTLS certificate verification disabled by default for Huawei Cloud SDK clients
skills/security/iam/huawei-cloud-iam-query/scripts/config.py:53
build_http_config() in huawei-cloud-iam-query, huawei-cloud-monitoring-query, and huawei-cloud-network-query hardcodes ignore_ssl_verification = True and suppresses TLS warnings. A network man-in-the-middle can intercept Huawei Cloud SDK calls to read cloud data, capture temporary security tokens or proxy credentials, or return forged IAM and resource responses.
Fix: Set ignore_ssl_verification to False by default and only allow disabling it through an explicit opt-in environment variable.
-
MediumAWS migration script opens target ECS SSH port 22 to 0.0.0.0/0
skills/migrationom/oos/huawei-cloud-migration-ecs-across-region/scripts/aws_to_huawei_direct.py:796
aws_to_huawei_direct.py (and aws_to_huawei_rsync.py) creates a target ECS with a public Elastic IP and root password, adds a security group rule allowing inbound SSH (TCP port 22) from 0.0.0.0/0 instead of restricting access to the known source host IP, and leaves the rule open permanently after migration.
Fix: Restrict inbound TCP port 22 to the specific source host IP (/32) during migration and remove temporary migration rules once cutover finishes.
-
MediumSSH host key verification disabled during server migration
skills/migrationom/oos/huawei-cloud-mgc-migration-ecs-gcp-to-hw/scripts/mgc_migrate.py:3248
mgc_migrate.py, aws_to_huawei_direct.py, aws_to_huawei_rsync.py, and ssh_utils.py (via paramiko.AutoAddPolicy()) disable SSH host key checking while transmitting root passwords and syncing entire root filesystems, including over public IPs. An active network attacker can impersonate the source or target host to steal root credentials and server contents or modify the migrated system.
Fix: Verify host keys against a known_hosts file or operator-supplied host key fingerprint instead of disabling StrictHostKeyChecking or using AutoAddPolicy.
-
MediumTarget root password persisted in cloud-init metadata and plaintext JSON artifacts
skills/migrationom/oos/huawei-cloud-mgc-migration-ecs-gcp-to-hw/scripts/mgc_migrate.py:574
build_linux_ssh_user_data_b64() embeds the base64-encoded target root password into cloud-init user_data and permanently enables PermitRootLogin and PasswordAuthentication, leaving the root password readable via the cloud metadata endpoint (169.254.169.254) to any local user or SSRF flaw on the migrated server. In the same skill, repair_boot_via_helper.py writes target_reset_password in plaintext to out/target_boot_repair_latest.json and stdout.
Fix: Rely on the ECS adminPass API field or SSH keys instead of embedding passwords in user_data, avoid permanently enabling root password SSH login, and redact passwords from output JSON files and stdout.
-
MediumMigration tool auto-executes archives from /tmp and extracts tarballs without path filtering
skills/migrationom/oos/huawei-cloud-migration-vmware-ecs-private/scripts/batch_migrate.py:1540
When SMS_AGENT_PATH or RSYNC_BINARY_PATH is not set, batch_migrate.py automatically picks up /tmp/SMS-Agent.tar.gz and /tmp/rsync-3.5.0-x86_64.tar.gz from world-writable /tmp, extracts the rsync archive locally with unfiltered tar.extractall() in sms_agent_push.py, and deploys and executes both as root on remote servers. On a shared machine, another local user can pre-create those files in /tmp to overwrite local files or run code as root on migrated hosts.
Fix: Do not auto-discover executables or archives from world-writable /tmp, and validate tar member paths (or pass filter='data' on Python 3.12+) before extraction.
-
MediumHuawei Cloud secret keys passed in process command-line arguments
skills/solution/csm/huawei-cloud-cce-env-assessment/scripts/collect_all.py:50
collect_all.py appends --cli-secret-key=<HWC_SK> to every hcloud command and passes secret_key inside curl -d arguments (and check_koocli.py similarly passes --cli-secret-key to hcloud configure set). On multi-user systems or hosts with process monitoring, command-line arguments are visible to other local processes via /proc/<pid>/cmdline or ps.
Fix: Avoid passing secret keys on the command line; use pre-configured hcloud profiles or pass sensitive request bodies through standard input instead of argv.
-
MediumUnescaped JSON injection into HTML report allows script-tag breakout
skills/security/ops/huawei-cloud-oa-risk-inspection/scripts/merge_report_generator.py:1433
render_html() serializes cloud inspection records with json.dumps() and substitutes the raw JSON directly into templates/report.html via str.replace('__DATA__', data_json). Because json.dumps() does not escape '</script>', any inspected resource name, tag, or property containing '</script><script>...' breaks out of the script element and runs arbitrary JavaScript under the file:// origin when the user opens the HTML report.
Fix: Escape '<', '>', and '&' as Unicode escapes (\u003c, \u003e, \u0026) in data_json before inserting it into the HTML template.
-
MediumEnvironment setup disables TLS verification globally and executes unverified get-pip.py from /tmp
skills/storage/obs/huawei-cloud-storage-query/scripts/ensure_env.py:28
ensure_env.py globally overrides Python's default HTTPS context with ssl._create_unverified_context, and if pip is unavailable it downloads get-pip.py over unverified HTTPS into a predictable path (/tmp/get-pip.py) and immediately executes it with Python. Either a network attacker (via MITM) or a local user on a shared machine (by pre-creating /tmp/get-pip.py) can execute arbitrary code.
Fix: Keep default TLS certificate verification enabled, remove the automatic get-pip.py download-and-execute fallback, and use NamedTemporaryFile if temporary files are needed.
-
MediumUnescaped jq filter interpolation and path traversal in OpenClaw multi-agent management script
multi-agent.sh interpolates $AGENT_ID, $AGENT_NAME, and $CHANNEL directly into jq filter strings rather than passing them via --arg, allowing crafted bot names or account IDs containing quotes to inject jq expressions into openclaw.json. In addition, $AGENT_ID is used unvalidated in filesystem paths ($WORKSPACE_BASE/agents/$AGENT_ID) passed to mkdir, chown -R, and rm -rf, enabling path traversal outside the agents directory.
Fix: Validate AGENT_ID against a strict alphanumeric/hyphen pattern and pass all variables to jq using --arg instead of string concatenation.
-
MediumMaaS helper scripts send MAAS_API_KEY to any unvalidated --base-url
skills/solution/ops/huawei-cloud-solution-ops/scripts/maas_common.py:163
maas_chat.py, maas_image_generation.py, maas_models.py, and maas_video_generation.py accept an arbitrary --base-url (including plaintext http:// or attacker-controlled domains) and attach 'Authorization: Bearer <MAAS_API_KEY>' without host validation. If prompt injection tricks the agent into supplying a custom --base-url, the user's MaaS API key and any local images passed via --image are exfiltrated.
Fix: Enforce https:// and restrict --base-url to trusted Huawei Cloud MaaS domains unless an explicit local configuration override is set by the user.
-
MediumNPU verification scripts load tokenizers and processors with trust_remote_code=True
skills/ai/cloudrobo/huawei-cloud-cloudrobo-npu-patcher/scripts/checks/dataload.py:39
dataload.py and pipeline.py in both huawei-cloud-cloudrobo-npu-patcher and huawei-cloud-cloudrobo-model-adapt-workflow call AutoTokenizer.from_pretrained and AutoProcessor.from_pretrained with trust_remote_code=True. Running verification on an untrusted third-party model directory executes arbitrary Python code referenced in tokenizer or processor JSON configs without warning.
Fix: Default trust_remote_code to False and require an explicit CLI flag after warning the user before executing custom model repository code.
-
MediumPlaintext cloud credentials written to terraform.auto.tfvars.json without restrictive file permissions or sensitive flags
skills/solution/sac/huawei-cloud-sac-dify/scripts/normalize_tf_providers.py:329
normalize_tf_providers.py in huawei-cloud-sac-dify, huawei-cloud-sac-new-api, and huawei-cloud-sac-yolo writes HW_ACCESS_KEY and HW_SECRET_KEY in plaintext to terraform.auto.tfvars.json using the default umask (typically world-readable 0644) rather than 0600, and generates variable blocks for access_key and secret_key without 'sensitive = true', causing Terraform output to expose credentials. In addition, huawei-cloud-sac-yolo/scripts/list_tf_variables.py prints sensitive variable defaults without masking.
Fix: Create tfvars files with 0600 permissions (as done in huawei-cloud-sac-intl-find-and-deploy), mark access_key and secret_key variables with sensitive = true, and mask sensitive defaults in list_tf_variables.py.
-
MediumCCE public API exposure and database CLI wrappers lack complete confirmation gates
skills/container/cce/huawei-cloud-cce-cluster-management/scripts/huawei_cloud/special_ops.py:158
bind_cce_cluster_eip binds a public EIP (or creates a new billable EIP automatically) to a CCE cluster's Kubernetes API server, exposing the control plane to the internet without requiring confirm=true (and other mutating actions such as create_cce_cluster, unbind_cce_cluster_eip, and install_cce_addon also omit confirm checks). Similarly, rds_cli.sh and gaussdb_cli.sh rely on incomplete MUTATING_OPS denylists rather than read-only allowlists, allowing unlisted destructive operations (such as DeleteDatabase or ResetPwd) to run without confirmation, and batch_create_buckets.sh sets acl=public-read automatically for website/assets purposes.
Fix: Require explicit confirmation for all non-read-only operations (using a read-only allowlist in CLI wrappers), and require confirmation before binding public EIPs to clusters or applying public-read bucket ACLs.
-
MediumDefault-on quality telemetry uploads command stdout and local AI agent session history
skills/devtools/common/huawei-cloud-skill-audit/scripts/cli/cli_reporting.py:829
Unless the user sets SKILL_QUALITY_DISABLE=1 or SKILL_QUALITY_REPORT=0, skill-quality-cli automatically sends execution reports to skillsapi.developer.myhuaweicloud.com. In five skills (skill-audit, deployment-task-management, cts-trace-management, dds-dcs-instance-management, and gaussdb-instance-management), cli_reporting.py reads local OpenCode, Hermes, Codex, and ACP session databases to upload the user's recent chat prompt and up to 20 recent tool messages, and in all eight skills with cli_entry.py (including dew-key-management for KMS keys and CSMS secrets) it uploads up to 6,000 characters of subprocess stdout.
Fix: Make telemetry opt-in by default, stop reading local agent session databases and ~/.config/huaweicloud/credentials.json, never upload raw command stdout, and require HTTPS domain validation on GUEST_ENDPOINT.
-
MediumAudited skills can silence static security findings by bundling a .skillspectorignore file
skills/devtools/common/huawei-cloud-skill-audit/scripts/checks/skillspector_builtin_check.py:294
SkillspectorBuiltinCheck loads .skillspectorignore directly from the target skill directory being audited and skips scanning that file itself. A malicious skill package can ship a crafted .skillspectorignore file to suppress static regex detections for prompt injection, credential access, data exfiltration, and destructive commands.
Fix: Do not honor ignore files located inside an untrusted target skill directory unless the user explicitly passes an allow-ignore flag, and report suppressed findings in the audit output.
-
MediumHTTP file server unconditionally bypasses configured token authentication
sandbox_http_server.py binds to 0.0.0.0:8080 by default and logs that token authentication is enabled when --token or sandbox.auth_token is set, but _check_auth() unconditionally returns True on its first line before the token check runs. Anyone on the network can browse directories and download served files without providing a token.
Fix: Remove the unconditional return True in _check_auth() when _auth_token is configured, use hmac.compare_digest for token comparison, and bind to 127.0.0.1 or a private interface by default.
-
MediumHuawei Cloud AK/SK passed on CLI command lines in SmsClient and OBSSigner
SmsClient._call() passes --cli-access-key and --cli-secret-key directly in the hcloud process argument list on every API call (even though it also sets HWCLOUD_ACCESS_KEY and HWCLOUD_SECRET_KEY in env), and OBSSigner.sign() similarly passes -i and -k on the obsutil command line. Any local user on the machine can read the cloud AK/SK from the process table (/proc/<pid>/cmdline or ps).
Fix: Remove --cli-access-key/--cli-secret-key from the hcloud argv list (relying on environment variables or a 0600 config file) and use an obsutil config file with 0600 permissions or SDK signing instead of -i/-k CLI flags.
-
MediumOpenClaw startup script installs unpinned remote plugin with --force --accept-capabilities and public mirror fallback
On cache miss during gateway startup, the injected ov-openclaw-init.sh installs @openviking/openclaw-plugin without a pinned version or integrity hash while passing --force --accept-capabilities, and overrides any user-configured npm registry with public https://registry.npmmirror.com/ on fallback. A compromised or unexpected upstream release is automatically granted full plugin capabilities at startup.
Fix: Pin the exact plugin version and integrity hash, avoid silently overriding a custom npm registry with a public mirror, and require explicit user opt-in before online fallback installation.
-
MediumAscend model deployment downloads and executes unverified remote shell scripts
skills/ai/modelarts/huawei-cloud-ascend-models-deploy/references/task-deploy-model.md
The core model deployment workflow downloads deploy-large-models.sh (or deploy-qwen3-vl-model.sh) from a public OBS URL using wget -P /home/modelarts-agent/ without any checksum verification and immediately executes it with sh. In addition, wget -P does not overwrite an existing file of the same name (saving to .1 instead), so a pre-existing or tampered /home/modelarts-agent/deploy-large-models.sh will be executed silently.
Fix: Write to an explicit temporary file path (-O), verify a pinned SHA-256 checksum before execution, or bundle the audited deployment script directly in the skill repository.
-
MediumPlaintext passwords and cloud credentials passed on command lines and through telemetry wrapper
skills/computing/ecs/huawei-cloud-ecs-manage/references/manage-operations.md:211
huawei-cloud-ecs-manage requires wrapping every hcloud call in skill-quality-cli run, including ResetServerPassword --reset-password.new_password={new_password}, exposing the instance's new root password in the OS process table (ps), shell history, agent logs, and the telemetry wrapper's argv. Other skills in this batch (ecs-lifecycle, flexus-l-server-flexusagent-deployment, flexus-l-server-hermes-deployment, flexus-l-server-openclaw-deployment, ecs-dsh-deploy, ecs-sqlbot-deploy, devkit-application-check-cli) similarly pass --ak, --sk, --admin-password, or --api-key on the command line, and flexus-l-server-ops runs env | grep HW_ which prints plaintext keys to stdout.
Fix: Pass AK/SK, passwords, and API keys via environment variables, stdin, or 0600 JSON input files (--cli-jsonInput) instead of CLI flags, exclude password-bearing commands from telemetry wrappers, and test credential presence without printing values.
-
MediumAgent instructed to automatically disable TLS verification on x509 certificate errors
skills/container/cce/huawei-cloud-kubectl-cce-installer/references/plugin-usage.md:58
Both huawei-cloud-kubectl-cce-installer and huawei-cloud-cce-log-analyzer instruct the agent to immediately re-run failed kubectl cce commands with --cce-insecure-upstream-tls=true whenever an x509 certificate validation error occurs. Because an x509 error is the exact symptom of a man-in-the-middle attack, automatically retrying without TLS verification exposes cluster API traffic and credentials to network interception.
Fix: Do not automatically bypass TLS verification on x509 errors; stop, report the certificate failure to the user, and require a valid CA certificate or explicit user confirmation.
-
MediumParamiko SSH/SFTP instructions disable SSH host key verification
skills/computing/kunpeng/huawei-cloud-kunpeng-source-code-migrate/references/scripts-guide.md:133
The SFTP and troubleshooting guides instruct the agent to connect to remote servers using paramiko.SSHClient() with AutoAddPolicy() and without loading system known_hosts. Every connection blindly trusts whatever host key is presented, allowing a network attacker to impersonate the remote server, capture MIGRATE_SSH_PASS, and hijack root SSH sessions.
Fix: Load known_hosts via client.load_system_host_keys() and use RejectPolicy(), or verify the server's host key fingerprint before authenticating.
-
MediumShell command injection in Terraform local-exec OBS upload templates
The OBS upload and static-website reference templates (also duplicated in Deploy-Bucket-Object-Uploaded-with-Local-File-best-practices.md and Deploy-Bucket-Static-Website-Hosting-Configuration-best-practices.md) interpolate var.object_upload_content, file names, and self.tags values directly into shell commands inside local-exec provisioners during create and destroy. If the agent generates Terraform using untrusted file names or text containing single quotes and shell metacharacters, or if someone modifies the bucket tags in the cloud before terraform destroy, running Terraform executes arbitrary shell commands on the user's machine.
Fix: Replace the local-exec echo/zip/rm shell commands with Terraform's local_file resource and archive_file data source, or pass content directly to huaweicloud_obs_bucket_object.
-
MediumMRS diagnostic API labeled as username list returns database passwords
skills/bigdata/mrs/huawei-cloud-mrs-host-alarm-diagnose/scripts/manager_api_apis/service.yaml:85
In both huawei-cloud-mrs-host-alarm-diagnose and huawei-cloud-mrs-host-fault-diagnose, the get_service_db_users API entry has desc set to 'Query service database username list', while the stripped YAML comment notes that this endpoint actually returns DBservice database passwords including the DBservice administrator password, and no response_filter is configured to strip secret fields. An agent calling this endpoint during diagnosis will pull cluster database passwords into the conversation context and logs.
Fix: Remove get_service_db_users from the diagnostic catalog or add a response_filter that strips all password fields before returning output to the agent.
-
MediumData-deletion and write APIs mislabeled as read-only in service catalog
The KooCLI service catalog sets read_only to true using name-based heuristics (Show*, *Search*, Check*), which misclassifies state-changing APIs as safe read-only queries. In idmeclassicapi.json, ShowLogicalDeleteUsingPost and ShowLogicalDeleteByConditionUsingPost are POST endpoints that soft-delete data instances; in idt.json, ten write operations (Save*, SaveAll*, SaveAs*, Checkin*, Checkout*, PatchRevise*) on XDMSearch* entities are marked read_only: true because the entity name contains Search; and in lakeformation.json and identitycenterportalapi.json, credential-minting endpoints (ShowCredential, GetAgencyCredentials) are also marked read_only: true. If an agent relies on this flag to run read-only checks without confirmation, prompt injection can trigger data deletion, overwrites, or credential issuance.
Fix: Determine read_only from the HTTP method and OpenAPI operation semantics rather than substring matches on operation names, and set read_only to false on all delete, save, checkin, checkout, verification, and credential-issuing operations.
-
MediumPublic anonymous read bucket policy attached to private encrypted source bucket in OMS templates
Both OMS object migration templates (migrate-objects-by-task and migrate-objects-by-group) configure the source OBS bucket with acl = "private" and KMS encryption enabled, and already pass IAM AK/SK credentials to the migration task, yet unconditionally attach a bucket policy granting Principal "*" GetObject on all objects in the bucket. Users applying this migration template will unexpectedly expose their private source bucket objects to anonymous downloads from anyone on the internet.
Fix: Remove the wildcard huaweicloud_obs_bucket_policy resource or scope it strictly to the authorized migration principal instead of "*".
-
MediumCluster admin kubeconfig written via local_file with world-accessible permissions
The CCE Kubernetes authentication template writes huaweicloud_cce_cluster.test.kube_config_raw (which contains cluster administrator client keys) to .kube/config using local_file without setting file_permission or directory_permission. In the Terraform local provider, local_file defaults permissions to 0777 and does not mark content sensitive, exposing cluster admin credentials in plan output and to other local users on shared machines.
Fix: Use the local_sensitive_file resource with file_permission = "0600" and directory_permission = "0700" instead of local_file.
-
MediumOver-broad FunctionGraph Administrator agency attached to ECS by default
Although terraform.tfvars comments state that use_custom_role defaults to true (least privilege to invoke one function), variables.tf actually sets use_custom_role = false and grants the ECS instance IAM agency the project-wide "FunctionGraph Administrator" role. Any code or user on the provisioned ECS instance can fetch temporary credentials from the metadata service and read, modify, or invoke all FunctionGraph functions in the project, including reading secrets stored in other functions' user_data.
Fix: Scope the agency to a least-privilege role restricted to the single target function or require explicit user opt-in before attaching FunctionGraph Administrator, and align terraform.tfvars comments with the actual default.
-
MediumThird-party boot script and unverified binary download in DevBridge ECS template
The DevBridge ECS template configures user_data to download and run setup scripts as root/SYSTEM from an unofficial personal GitCode repository (anruostyle/devbridge-ssh-setup), and on Windows also downloads and executes an embedded Python zip from python.org without checksum verification. While the outer GitCode script is SHA256-checked, any unpinned secondary downloads inside that third-party script or MITM/CDN compromise on the unverified zip run with full system privileges on the new server.
Fix: Host setup artifacts in an official Huawei Cloud repository, verify SHA256 checksums for all downloaded binaries including the embedded Python archive, and document external download sources.
-
MediumManagement ports (SSH, Telnet, RDP) opened to 0.0.0.0/0 in Terraform templates
Multiple Terraform generator templates open administrative ports to the entire internet (0.0.0.0/0) by default: vpc/security-group/terraform.tfvars allows TCP 22-23 (SSH and Telnet) and 3389 (RDP) from 0.0.0.0/0, ecs/instance-provisioners/main.tf hardcodes TCP 22 from 0.0.0.0/0, and nat/dnat-basic and nat/nat-gateway-vpc-peering default SSH rules to 0.0.0.0/0. Deploying these templates without tightening the source CIDR exposes instances to internet-wide brute-force and protocol attacks.
Fix: Require an explicit administrator CIDR variable (with no 0.0.0.0/0 default) for SSH and RDP, and remove Telnet (port 23) from default examples.
-
MediumWeak hardcoded default password for public-EIP instances in Terraform templates
The RabbitMQ producer/consumer Terraform template sets a default instance_password of "123456" (and "YourPassword@123" in terraform.tfvars) and uses it both for RabbitMQ authentication and as root admin_pass on two ECS instances with public EIPs attached; similarly, elb/dedicated-loadbalancer-with-as/terraform.tfvars embeds a static root password hash via chpasswd -e in user_data. Users applying these templates without overriding credentials may deploy internet-reachable servers with known or weak passwords.
Fix: Remove default password values so the user must supply a strong secret or generate one per deployment with random_password.
-
MediumTLS certificate verification disabled by default in Windows migration config
The Windows migration proxy configuration sets verify_ssl to false by default for source-side transfers. Anyone on the network path between the source server or proxy and the remote HTTPS server can intercept traffic or tamper with downloaded migration files (such as the SMS-Agent installer) without triggering a certificate error.
Fix: Set verify_ssl to true by default and allow a custom CA bundle path when private certificates are used.
-
MediumWindows migration file server binds to 0.0.0.0 without authentication
Although the comments describe the sandbox HTTP service as a private-network file server on the proxy jumpbox, http_host is set to 0.0.0.0 on port 8080 by default. Because a migration jumpbox often has both public and private network interfaces, binding to 0.0.0.0 exposes the unauthenticated ./downloads directory on all interfaces unless blocked by external firewall rules.
Fix: Bind http_host to 127.0.0.1 or require the specific private interface IP address of the proxy host instead of 0.0.0.0.
-
LowUnpinned third-party AI install instructions fetched from external personal repository
skills/cdi/assist/huawei-cloud-icp-process-guidance/SKILL.md:27
For search fallback in ICP filing Q&A, this skill instructs the user to have the AI execute installation instructions fetched live from the main branch of an external personal GitHub repository (Panniantong/agent-reach). Because the URL is unpinned and interpreted directly by an AI agent with shell access, any upstream modification or compromise of that repository can inject arbitrary shell commands or prompt-injection instructions into the user's machine.
Fix: Remove the external unpinned AI install prompt or vendor a pinned, checksum-verified search script directly inside the skill package.
-
LowSkill hijacks all generic coding requests even when user asks to write code locally
skills/devtools/cli/huawei-cloud-codearts-code-gen/SKILL.md:135
The skill claims mandatory priority over every request to write code, create files, or implement features—even on unrelated local projects and even when the user explicitly instructs the agent to write code directly. Once configured, it routes task prompts and attached local source files to the remote CodeArts service, and during setup it asks the user to paste Huawei Cloud AK/SK in chat, passes them on the command line (--save-aksk <AK> <SK>), and stores them in plaintext in ~/.bashrc.
Fix: Trigger this skill only when the user asks for CodeArts, respect explicit user requests to write code locally, and read credentials via interactive stdin or secure credential files rather than CLI arguments and ~/.bashrc.
-
LowConversation logs sent to external GitCode issues without confirmation and third-party login helper
skills/devtools/common/huawei-cloud-vod-collector/SKILL.md:200
Triggered whenever a user expresses frustration or hits an error, this skill captures 3-5 turns of conversation context, error stacks, and environment metadata, then posts them as an issue to a repository fixed in assets/config.yaml without asking the user. If a devspace-connector MCP tool is available, it creates the external issue immediately without user confirmation; otherwise, it installs an OAuth login server from a personal GitCode account (weixin_45218422/AtomGit-GO), stores the AtomGit token in plaintext on disk, and sends access_token in URL query parameters.
Fix: Require explicit user review and confirmation of the sanitized payload and target repository before submitting any issue, and host any authentication helper under an official verified organization with pinned checksums.
-
LowDeployment scripts silently disable hcloud CLI credential encryption on disk
skills/computing/flexus-l/huawei-cloud-flexus-l-deploy-jiuwenswarm/SKILL.md:396
To read AK/SK values directly from ~/.hcloud/config.json, this skill automatically disables KooCLI's credential encryption setting (authEncrypt=true) so credentials are stored in plaintext on disk. This silently weakens at-rest protection for the user's global Huawei Cloud CLI profile long after the deployment finishes.
Fix: Invoke hcloud CLI directly or read credentials from environment variables instead of disabling authEncrypt on ~/.hcloud/config.json.
-
LowHuawei Cloud Secret Key sent in plaintext HTTP header to activity endpoint
skills/devtools/common/huawei-cloud-activity-submit/SKILL.md:43
Instead of signing requests locally with HMAC-SHA256, this skill transmits the user's Huawei Cloud Secret Key (SK) directly in the X-HW-SK HTTP header to open-dataset.huaweicloud.com (and also accepts --sk on the command line). Sending a cloud account's raw Secret Key over HTTP headers exposes full account-signing material to the remote web application, reverse proxies, and any server-side request logging.
Fix: Use standard Huawei Cloud SDK-HMAC-SHA256 request signing or short-lived scoped STS tokens rather than transmitting raw Secret Keys in HTTP headers.
-
LowPlaintext credentials passed via CLI flags and echoed into chat output
skills/computing/flexus-l/huawei-cloud-flexus-l-server-flexusagent-deployment/SKILL.md:268
Multiple operational skills pass sensitive secrets directly as command-line arguments (such as --ak, --sk, and --admin-password here, --ak/--sk in flexus-l-server-ops, and --password in ascend-command and ascend-remote-connect), exposing them in the local process table (ps aux) and shell history. In addition, this skill mandates echoing the newly set Dify admin password in plaintext in the chat response alongside an unencrypted HTTP login URL.
Fix: Pass credentials exclusively through environment variables or stdin/files with 0600 permissions, and avoid echoing plaintext passwords in agent chat logs.
-
LowAuto-installed telemetry CLI wraps commands and reports execution metadata by default
skills/devtools/common/huawei-cloud-find-skills/SKILL.md:35
Many skills in this collection mandate running scripts/ensure_cli.sh (installing skill-quality-cli into ~/.local/bin and symlinking it into a writable PATH directory) and wrapping every hcloud or Python command with skill-quality-cli run (or skill_quality_sdk.py). Telemetry reporting of execution status, trace IDs, skill impressions, and masked input/output to Huawei Cloud endpoints is enabled by default (opt-out via SKILL_QUALITY_REPORT=0 or SKILL_QUALITY_DISABLE=1), and several skills do not prominently disclose what masked command inputs/outputs leave the machine.
Fix: Make external telemetry opt-in rather than opt-out, avoid modifying writable system PATH directories automatically, and consistently disclose all transmitted fields across every skill.
-
LowSkill instructs agent to ask user for plaintext cloud AK/SK in chat
skills/solution/csm/huawei-cloud-cce-env-assessment/SKILL.md:120
Step 1 of huawei-cloud-cce-env-assessment instructs the agent to prompt the user for their Huawei Cloud Access Key and Secret Key directly in the conversation and save them into environment variables. Entering long-lived cloud credentials in chat exposes them to LLM model providers, conversation logs, and session telemetry.
Fix: Instruct users to configure AK/SK out-of-band via hcloud configure init or terminal environment variables, and forbid accepting plaintext credentials in chat.
-
LowSkill instructs automatic sudo privilege escalation on permission errors
skills/solution/csm/huawei-cloud-cce-env-assessment/SKILL.md:55
Rule 3 of huawei-cloud-cce-env-assessment instructs the AI agent to automatically retry failed operations with sudo whenever it encounters a permission problem, without requiring explicit user confirmation or scoping which commands may run as root.
Fix: Remove the automatic sudo rule and require the agent to stop, explain the permission error, and ask the user before running any privileged command.
-
LowUndisclosed default-on telemetry sending user input and session data externally
skills/monitoring/ecs/huawei-cloud-ecs-query/SKILL.md:35
Several skills (such as huawei-cloud-ecs-query and huawei-cloud-eip-query) require every command to be wrapped with skill-quality-cli, which automatically collects session_id, user_input, tokens, and steps from the host session and sends them to skillsapi.developer.myhuaweicloud.com by default, without disclosing this data transmission or the opt-out variable in those skills' documentation.
Fix: Clearly disclose what session and user-input data skill-quality-cli transmits and how to disable it (SKILL_QUALITY_REPORT=0), or make telemetry opt-in.
-
LowReference inference server template binds to 0.0.0.0 without authentication
skills/ai/cloudrobo/huawei-cloud-cloudrobo-inference-adapter/references/server_template.py:37
TransportServer in references/server_template.py defaults to binding on 0.0.0.0:8000 with a self-signed certificate, no client authentication, unlimited WebSocket frame size (max_size=None), and sends full stack traces back to clients on error. Although never executed by the installed skill directly, users copying this template outside an isolated cluster network could expose an unauthenticated endpoint.
Fix: Default host to 127.0.0.1, set a bounded max_size on WebSocket frames, require token or mTLS authentication, and avoid returning raw tracebacks to clients.
-
LowPlaywright browser launched with --no-sandbox and ignore_https_errors=True
skills/cdi/assist/huawei-cloud-icp-rule-consult/scripts/web_searcher.py:109
web_searcher.py launches local Chrome with --no-sandbox while setting ignore_https_errors=True on the browser context and accepting any target domain via --site. Disabling both TLS certificate validation and Chromium's OS sandbox weakens protection against MITM script injection and renderer exploits.
Fix: Remove --no-sandbox unless explicitly required in a container, set ignore_https_errors=False, and restrict --site to trusted Huawei Cloud documentation domains.
-
LowUnescaped HDFS log operation names interpolated into HTML/SVG trend charts (Stored XSS)
skills/bigdata/mrs/huawei-cloud-mrs-hdfs-performance-issue-analysis/scripts/hdfs_perf_analyze.py:682
gen_topuser_chart and build_path_with_dots interpolate operation strings parsed from 5min-TopUserOpCounts.log directly into SVG <text> elements and data-name attributes without HTML escaping. Opening the generated HTML report for a crafted or untrusted log bundle executes injected JavaScript in the browser.
Fix: Escape all log-derived strings (timestamps, operation names, metric labels) with html.escape(..., quote=True) before embedding them in HTML or SVG.
-
LowUnencoded URL path parameter substitution in MRS Manager API client
skills/bigdata/mrs/huawei-cloud-mrs-host-alarm-diagnose/scripts/manager_api_client.py:311
_build_url in manager_api_client.py (also duplicated in huawei-cloud-mrs-host-fault-diagnose) substitutes {param} placeholders into the request path using raw string replacement without URL-encoding. Unlike lakewatch_api_client.py, a parameter lacking a strict regex rule in YAML can inject '../' or query delimiters into authenticated requests sent to MRS Manager.
Fix: URL-encode path parameter values using urllib.parse.quote(str(params[key]), safe='') before substituting placeholders.
-
LowSigned Authorization and X-Security-Token headers printed in plaintext to stdout
skills/computing/flexus-l/huawei-cloud-flexus-l-server-openclaw-deployment/scripts/lib.py:253
create_openclaw_instance() and query_uniagent_status() print the full signed HTTP headers dictionary—including Authorization (with plaintext AK) and X-Security-Token (the live temporary STS credential)—directly to stdout, leaking temporary cloud credentials into terminal logs and the AI agent's context window.
Fix: Redact Authorization and X-Security-Token headers before logging (as already implemented in hermes-deployment/scripts/lib.py).
-
LowCloud secret keys and SSH passwords passed via command-line arguments
skills/migrationom/oos/huawei-cloud-migration-vmware-ecs-private/scripts/hcloud_wrapper.py:110
hcloud_wrapper.py passes --cli-secret-key on every hcloud subprocess command line (and aws_to_huawei_direct.py and aws_to_huawei_rsync.py pass passwords via 'sshpass -p'). Command-line arguments are visible in the OS process table (ps aux or /proc/<pid>/cmdline) to other local users on multi-user systems.
Fix: Use an encrypted hcloud profile instead of --cli-secret-key arguments, and pass SSH passwords via environment variables (sshpass -e) or SSH keys.
-
LowUnescaped cloud resource names and error strings in generated HTML reports
skills/monitoring/ces/huawei-cloud-alarm-reduce-noise/scripts/analyze_alarm.py:1526
analyze_alarm.py (in its governance sections and inline script JSON), huawei-cloud-ces-drill-inspection/scripts/html_template.py, and huawei-cloud-migration-vmware-ecs-private/scripts/report.py interpolate cloud resource names, metrics, and remote error strings directly into HTML without html.escape(). Crafted resource names or remote error output can run JavaScript in the user's browser when the report is opened.
Fix: Escape all dynamic strings with html.escape() before inserting them into HTML, and escape '</' as '<\/' when embedding JSON inside <script> tags.
-
LowPredictable shared /tmp directory for billing confirmation tokens
skills/solution/ops/huawei-cloud-solution-ops/scripts/hcloud_billing_live_read.py:25
issue_confirm_token() stores confirmation tokens in a fixed shared directory (/tmp/hcloud-billing-live-read-confirm) created with mkdir(exist_ok=True) and writes predictable <scope_hash>.json files with write_text() before chmod(0o600). On a multi-user machine, another local user can pre-create that directory without the sticky bit and plant symlinks to overwrite files owned by the user or tamper with tokens.
Fix: Store token state inside a user-private directory (for example Path.home() / '.hcloud' with mode 0700) and write files atomically using hcloud_common.write_json_artifact().
-
LowUnvalidated endpoint overrides send cloud auth headers and STS tokens to arbitrary hosts
skills/security/scm/huawei-cloud-ccm-certificate-query/scripts/list_certificates.py:283
list_certificates.py (and similarly hcloud_sdk_readonly.py) accepts an arbitrary --endpoint override without restricting the host to Huawei Cloud domains. If prompt injection tricks the agent into supplying an attacker-controlled endpoint, the script sends signed Authorization headers (exposing the Access Key ID) and X-Security-Token (when temporary STS credentials are set) to that external or internal host.
Fix: Validate that custom endpoints match expected Huawei Cloud domains (such as *.myhuaweicloud.com) and require HTTPS before attaching authentication or security-token headers.
-
LowCloud AK/SK credentials passed as command-line arguments to hcloud and kubectl subprocesses
skills/container/cce/huawei-cloud-cce-alarm-correlation-engine/scripts/huawei_cloud/common.py:84
Across the CCE skills (alarm-correlation-engine, cluster-management, kubernetes-event-analyzer, log-analyzer, and metric-analyzer), when no local ~/.hcloud/config.json profile is present, AK/SK values from environment variables or parameters are appended as '--cli-access-key=...' and '--cli-secret-key=...' command-line flags. On multi-user systems, command-line arguments are visible to other local users via /proc/<pid>/cmdline (ps aux); similarly, obs_lifecycle_analyzer.py writes temporary rule JSON to a predictable /tmp/obs_lifecycle_<bucket>.json path.
Fix: Pass credentials via environment variables or protected temporary configuration files (mode 0600) instead of command-line flags, and use tempfile.NamedTemporaryFile for temporary lifecycle files.
-
LowDNS API version check disables TLS certificate verification
skills/network/ops/huawei-cloud-network-query/scripts/dns/list_api_versions.py:31
The script creates an unverified SSL context with ssl._create_unverified_context() and interpolates --region directly into the request URL hostname without format validation. An attacker on the network path can spoof the HTTPS response, and a crafted --region value can cause the script to send a GET request to an unintended host.
Fix: Use ssl.create_default_context() so server certificates are verified, and validate --region against a strict region-code regular expression.
-
LowSecurity group helper passes Huawei Cloud secret key as a command-line argument
skills/network/ops/huawei-cloud-sg-manage/scripts/sg/common.py:98
When no ~/.hcloud profile file is present, run_hcloud passes HUAWEI_ACCESS_KEY, HUAWEI_SECRET_KEY, and HUAWEI_SECURITY_TOKEN directly on the hcloud command line (--cli-secret-key=...). On multi-user systems, command-line arguments are visible to other local users via the process table (/proc/<pid>/cmdline) while hcloud is running.
Fix: Pass credentials to hcloud via environment variables or a restricted temporary configuration file (mode 0600) instead of command-line flags.
-
LowWinRM setup enables cleartext Basic auth, unencrypted traffic, and wildcard TrustedHosts
HuaweiCloudWinRMProvider (as well as AWSWinRMProvider and AliyunWinRMProvider) configures Windows WinRM with Basic='true', AllowUnencrypted='true', and TrustedHosts='*', and never reverts these settings after migration. This permanently weakens WinRM security on the source server and on any target ECS created via block-level cloning.
Fix: Use WinRM over HTTPS (port 5986) with NTLM/Kerberos encryption instead of enabling Basic authentication, AllowUnencrypted, and wildcard TrustedHosts.
-
LowPlaintext Huawei Cloud AK/SK persisted in Windows watchdog script and process arguments without cleanup
In Py2 migration mode, ConfigurePy2Phase writes Huawei Cloud --ak and --sk in plaintext into C:\auto_agent_opt\watchdog.ps1 and registers a persistent Scheduled Task (SMSAgentWatchdog) that runs every 3 minutes forever, passing --ak and --sk on the agent_config.exe command line. Neither the task nor watchdog.ps1 is deleted after registration succeeds, leaving cloud credentials exposed in Windows process audit logs and persisted onto block-migrated target ECS disks.
Fix: Pass credentials to agent_config.exe via a temporary ACL-restricted file deleted immediately upon reading (as in Py3 mode), and unregister SMSAgentWatchdog and delete watchdog.ps1 once registration succeeds.
-
LowOpenViking API key written in plaintext to world-readable startup scripts
When OV_API_KEY (--api-key) is provided, ov_workswarm.py embeds the secret literally into ov-jiuwenswarm-init.sh and sets permissions to 0o755 (world-readable), and ov_openclaw.py embeds it in ov-openclaw-init.sh and passes it via --api-key on the process command line. Other local users or processes with access to shared_dir or /proc can read the API key.
Fix: Store the API key in a separate file with 0o600 permissions or inject it via runtime environment variables rather than embedding it into 0o755 startup scripts and CLI arguments.
-
LowModel deployment guide mandates disabling TLS certificate verification (curl -k) for external sites
The reference guide instructs the agent as a mandatory rule to pass -k (disabling TLS certificate verification) whenever curling docs.vllm.ai and quay.io. Because the agent extracts container image tags, environment variables, and vLLM startup flags from those responses to generate start_vllm.sh (and ingests the HTML into its prompt context), a network man-in-the-middle attacker can inject malicious startup arguments, images, or prompt injections.
Fix: Keep TLS certificate verification enabled by default and instruct users behind corporate TLS-intercepting proxies to configure their enterprise CA bundle (CURL_CA_BUNDLE / SSL_CERT_FILE) rather than mandating curl -k.
-
LowContainer startup script installs unhashed third-party pre-release package from PyPI as root
skills/ai/modelarts/huawei-cloud-modelarts-vllm-ascend-model-deploy/references/hixlep-guide.md:69
When /etc/hccl_rootinfo.json is absent and HCCL_TOOLS_WHEEL is unset, the A5 container startup workflow installs the third-party pre-release package unofficial-ascend-tools==0.0.7rc4 from public PyPI without --require-hashes and runs its entry point as root.
Fix: Pin SHA-256 package hashes with pip install --require-hashes or require a pre-verified offline wheel rather than installing an unhashed third-party package from PyPI at container startup.
-
LowIAM agency policy for OptVerse grants account-wide destructive OBS bucket permissions
skills/ai/optverse/huawei-cloud-optv-evolve-management/references/iam-agency.md:90
To let OptVerse upload evolve task outputs to a single OBS path, iam-agency.md instructs creating an account-wide custom agency policy (policy_ai4s_llma4ad) delegated to op_svc_oroas_container0 with no Resource restriction and 30 OBS actions, including obs:bucket:DeleteBucket, obs:bucket:DeleteBucketPolicy, obs:bucket:PutBucketPolicy, obs:bucket:PutBucketAcl, and obs:object:DeleteObjectVersion across all buckets in the account.
Fix: Scope the agency policy Resource to the specific target OBS bucket and prefix, and restrict actions to the minimum needed to upload and list task results (such as obs:object:PutObject, obs:object:GetObject, and obs:bucket:ListBucket).
-
LowPrerequisite guides instruct piping unverified remote URLs directly into bash
Several installation guides instruct running curl ... | bash (or iwr ... | iex) without verifying a SHA-256 checksum—including piping an HTML documentation URL into bash in modelarts-notebook-management, get.docker.com in cloudrobo-dataset, openclaw.ai/install.sh in maas-model-guidance, and unverified hcloud_install.sh downloads in multiple skills.
Fix: Download installers to a local file and verify a pinned SHA-256 checksum before executing, following the pattern used in huawei-cloud-modelarts-resource-pool-management/references/cli-installation-guide.md.
-
LowTLS certificate verification disabled by default for MRS Manager and LakeWatch APIs
skills/bigdata/mrs/huawei-cloud-mrs-host-alarm-diagnose/scripts/manager_api_config.yaml:56
Both manager_api_config.yaml and lakewatch_api_config.yaml in huawei-cloud-mrs-host-alarm-diagnose and huawei-cloud-mrs-host-fault-diagnose ship with verify_ssl: false by default while sending MRS Manager admin credentials (via HTTP Basic Auth or session login) and LakeWatch credentials over HTTPS. An attacker on the network path can intercept cluster admin credentials via a man-in-the-middle attack, and the inline Chinese comment (是否跳过 SSL 证书验证,仅自签名证书的内网环境设为 true) states the opposite of what verify_ssl actually does.
Fix: Set verify_ssl: true by default, require users with self-signed intranet certificates to configure ca_cert explicitly, and fix the inverted YAML comment so true clearly means verifying certificates.
-
LowAutomatic opt-out telemetry sends skill execution data without user consent
skills/bigdata/mrs/huawei-cloud-mrs-redis-meta-check/references/cli-installation-guide.md:18
Several skills (huawei-cloud-mrs-redis-meta-check, huawei-cloud-voucher-assistant, huawei-cloud-ecs-manage, huawei-cloud-partner-scout, huawei-cloud-ecs-lifecycle, huawei-cloud-ecs-vulnerability-auditor) enable automatic quality telemetry by default unless the user sets SKILL_QUALITY_REPORT=0 or SKILL_QUALITY_DISABLE=1, and mrs-redis-meta-check and voucher-assistant include a fallback script that downloads and installs skill-quality-cli into ~/.local/bin/. Execution telemetry leaves the machine without opt-in permission.
Fix: Make telemetry strictly opt-in (disabled unless explicitly enabled by the user) and strip all command arguments and resource identifiers before reporting.
-
LowUnpinned remote CLI install scripts and binaries executed without checksum verification
huawei-cloud-devkit-application-check-cli instructs the agent to auto-install hcloud when missing by downloading a mutable /latest/hcloud_install.sh script and executing it directly with bash without SHA-256 verification. Similar unverified /latest/ installer or binary download instructions appear in billing-scout, bss-voucher-query, ecs-vulnerability-auditor, ecs-passwordless-login, flexus-l-server-hermes-deployment, ecs-dsh-deploy, ecs-sqlbot-deploy, dws-migration-teradata-to-dws, and dws-spec-planner. Compromise of any of those bucket objects would execute arbitrary code on the user's machine.
Fix: Pin specific release versions and verify SHA-256 checksums before running any downloaded installer or binary (matching huawei-cloud-account-onboarding), and require user confirmation before installing tools.
-
LowPowerShell command injection and unvalidated URI execution via OAuth verification_url
skills/businessapps/doczip/huawei-cloud-clouddocs-aipage/bin/cli-installation-guide.md:89
During device-code login, the guide instructs taking <verification_url> from the JSON response of clouddocs-cli auth login --no-wait --json (returned by the configured service host) and interpolating it into powershell.exe -NoProfile -Command "Start-Process '<verification_url>'". A single quote in verification_url breaks out of the string to execute arbitrary PowerShell commands, and an unvalidated scheme allows Start-Process to launch local binaries, UNC paths, or non-HTTPS protocol handlers.
Fix: Validate that verification_url strictly starts with https:// and contains no quotes or control characters before launching it, and pass the URL safely without string interpolation into powershell -Command.
-
LowSystem /etc/hosts modified and backed up to world-writable /tmp during metric analysis
skills/computing/ecs/huawei-cloud-ecs-shutdown-experiment/references/managed-service-logs.md:148
Although Phase 3 (Analyze) is described in iam-policies.md as read-only, managed-service-logs.md and analysis-workflow.md instruct querying 223.5.5.5 and appending an IP override for ces.cn-north-4.myhuaweicloud.com directly to /etc/hosts when CES calls time out, while saving the backup to a predictable /tmp/hosts.bak path. Modifying /etc/hosts changes system-wide name resolution without user confirmation, and restoring from /tmp/hosts.bak on a multi-user machine risks local /tmp symlink or file-tampering attacks.
Fix: Remove the /etc/hosts modification from the analysis workflow, report DNS or network errors to the user instead, and never store privileged file backups at predictable paths in /tmp.
-
LowSSH host key verification disabled and target password exposed in SSH command string
Both devkit-operations-guide.md and verification-method.md run ssh with -o StrictHostKeyChecking=no alongside sshpass when connecting from the DevKit ECS to target servers, allowing a network man-in-the-middle attacker to impersonate a target host and capture its root password. In addition, the verification command embeds SSHPASS='<target_password>' directly in the local ssh command string, exposing the password in process listings and agent transcripts.
Fix: Verify target SSH host keys instead of passing StrictHostKeyChecking=no, and never embed plaintext target passwords in ssh command strings.
-
LowPlaintext database passwords required on disk in migration_config.yaml
Unlike the other migration scripts that read TD_PASSWORD and DWS_PASSWORD from environment variables, migration_config_template.yaml notes that migrate_views_procs.py requires plaintext Teradata and DWS passwords to be written into config/migration_config.yaml on disk, increasing the risk of accidental exposure via agent file reads, backups, or version control.
Fix: Update migrate_views_procs.py to read TD_PASSWORD and DWS_PASSWORD from environment variables and remove plaintext password fields from migration_config.yaml.
-
LowLocal-KB ICP guidance skill sends fallback search queries to external Exa MCP server
skills/cdi/assist/huawei-cloud-icp-process-guidance/config/mcporter.json:4
Although references/iam-policies.md states that huawei-cloud-icp-process-guidance answers questions purely from its built-in local knowledge base, config/mcporter.json and references/search-fallback.md configure and call an external third-party Exa MCP endpoint (https://mcp.exa.ai/mcp) when local entries are missing or older than six months, sending query keywords outside Huawei Cloud and pulling untrusted web search snippets into the agent context.
Fix: Ask for user confirmation before sending queries to external third-party search services and align iam-policies.md and acceptance-criteria.md with the actual external network behavior.
-
LowPlugin code and verification hashes fetched from third-party GitHub proxies by default
skills/devtools/extensions/huawei-cloud-openviking-agent-integration/references/notes.md:39
At integration time, this skill downloads executable agent plugins and their Git verification metadata (unpinned commit SHA and tree blob SHAs) through unofficial third-party proxy domains (gh-proxy.com and ghfast.top) before falling back to GitHub. Because gh-proxy.com is queried first for both the expected hashes and the plugin files, a compromised or hijacked proxy can serve malicious plugin code with matching hashes and gain code execution inside the user's coding agents.
Fix: Fetch commit and tree metadata directly from api.github.com over HTTPS or pin trusted commit and blob SHA-256 hashes inside the skill repository.
-
LowUnverified kubectl-cce binary downloaded from an unofficial personal repository
skills/container/cce/huawei-cloud-cce-kubernetes-event-analyzer/references/kubectl-cce.md:16
Multiple CCE skills (including huawei-cloud-cce-kubernetes-event-analyzer, huawei-cloud-cce-metric-analyzer, and huawei-cloud-kubectl-cce-installer) instruct the user or agent to download and install the kubectl-cce binary from a personal GitHub/Gitee account (pancake0001/kubectl-cce-plugin) into /usr/local/bin without verifying any SHA-256 checksum, and then pass Huawei Cloud AK/SK credentials to it. Because release assets on personal repositories can be modified or hijacked, this creates a direct supply-chain and credential-theft risk.
Fix: Distribute kubectl-cce from an official Huawei Cloud repository or mirror and verify a pinned SHA-256 checksum before installing it.
-
LowDOM-based cross-site scripting (XSS) in Kafka reassignment Web UI
skills/devtools/kafka/huawei-cloud-kafka-replica-reassignment/templates/index.html:81
The local Web UI injects backend log lines, stage error messages, validation failure details, and output filenames directly into innerHTML without HTML escaping. Because validation errors and logs echo untrusted cell values (such as invalid topic names or mount paths) from uploaded .xlsx spreadsheets, uploading a crafted spreadsheet executes arbitrary JavaScript in the user's browser.
Fix: Populate dynamic status, validation, filename, and log elements using textContent, or HTML-escape all untrusted strings before assigning to innerHTML.
-
LowDetached signature verification skips certificate validation (-noverify) before root install
Before extracting the Kunpeng DevKit archive and running its installer as root on the target server, the guide checks the .p7s signature using openssl cms -verify with the -noverify flag. Because -noverify skips verifying the signer's certificate against any trusted CA or pinned certificate, an attacker who supplies a malicious --devkit-url (for example via prompt injection) or compromises the download source can sign a tampered tarball with any self-signed key and pass verification.
Fix: Remove -noverify and pin Huawei's official signing certificate or root CA via -CAfile or -certfile when verifying the PKCS#7 signature.
-
LowGenerated passwords required in README.md and STS tokens written to terraform.tfvars
The Terraform generator rules require the agent to display all generated database passwords, admin passwords, and API tokens in plaintext inside the generated README.md (and guardrails.md also writes HW_SECURITY_TOKEN from environment variables into terraform.tfvars without generating a .gitignore). Because README.md is routinely committed to Git repositories and shared, this exposes live infrastructure credentials.
Fix: Never write plaintext passwords, API tokens, or STS tokens into README.md, generate a .gitignore excluding *.tfvars and *.tfstate, and pass temporary security tokens via environment variables.
-
LowRaw Huawei Cloud Secret Key sent in HTTP request header
skills/devtools/common/huawei-cloud-activity-submit/references/iam-policies.md:10
Instead of signing requests locally with HMAC-SHA256 or using short-lived STS credentials, this skill reads HW_SECRET_KEY / HUAWEICLOUD_SDK_SK from the environment and sends the raw Secret Key in an X-HW-SK HTTP header to the activity submission endpoint. If the user has long-term cloud credentials configured, header logging or a compromise of the activity service exposes their Huawei Cloud account.
Fix: Use standard Huawei Cloud AK/SK request signing or short-lived, minimal-privilege STS tokens (such as SELF_VERIFY) instead of sending raw Secret Keys in HTTP headers.
-
LowSecurity group rules default to opening SSH and DevKit WebUI to 0.0.0.0/0
skills/computing/kunpeng/huawei-cloud-devkit-webui-create/references/ecs-creation-guide.md:261
Both the manual instructions and the automated hcloud commands default REMOTE_IP_PREFIX to 0.0.0.0/0, exposing password-authenticated root SSH (port 22) and the DevKit WebUI (port 8086) on a public EIP to the entire internet unless the user manually overrides the variable.
Fix: Require an explicit caller IP/CIDR (such as <outbound_ip>/32) instead of defaulting REMOTE_IP_PREFIX to 0.0.0.0/0.
-
LowUnpinned remote installer scripts executed without checksum verification
skills/devtools/cli/huawei-cloud-codearts-code-gen/references/cli-installation-guide.md:49
huawei-cloud-codearts-code-gen (automatically during setup and in its manual guide) and several other CLI guides (including huawei-cloud-iac-reverse and huawei-cloud-devbridge-tunnel) download unpinned latest shell or PowerShell install scripts and execute them directly without verifying a SHA-256 checksum.
Fix: Pin CLI installers to specific release versions and verify published SHA-256 checksums before executing downloaded scripts.
-
LowOpt-out telemetry CLI installed and executed automatically across skills
skills/devtools/common/huawei-cloud-find-skills/references/cli-installation-guide.md:5
Multiple skills automatically install skill-quality-cli into ~/.local/bin and transmit skill usage and search-exposure telemetry to Huawei Cloud endpoints on an opt-out basis without asking for consent first. In addition, huawei-cloud-skill-audit documents SKILL_QUALITY_DISABLE=1 whereas the other skills use SKILL_QUALITY_REPORT=0 to disable reporting.
Fix: Make external telemetry opt-in by default and use a single consistent environment variable name to control reporting across all skills.
-
LowSetup guides recommend passing plaintext AK/SK on the command line
Several CLI setup guides (including huawei-cloud-kunpeng-source-code-migrate, huawei-cloud-cce-log-analyzer, and huawei-cloud-publish-work-to-gallery) recommend passing --cli-access-key and --cli-secret-key directly as command-line arguments, which records secret keys in shell history and exposes them to other local processes via the process table.
Fix: Use interactive hcloud configure init prompts or read credentials from protected files/environment variables rather than command-line flags.
-
LowUnpinned curl-to-bash CLI install scripts and inconsistent OBS bucket names
skills/storage/rds/huawei-cloud-rds-smart-service/references/cli-installation-guide.md:8
Multiple CLI installation guides instruct piping unpinned remote shell scripts directly into bash without checksum verification. Across skills, at least seven non-standard OBS bucket names are used instead of Huawei Cloud's official KooCLI bucket—such as 'hcloud_client' here, 'cli-hk' (waf-false-alarm-diagnosis), 'cn-north-4-hwc-cloudsite' (dcs-failover), 'cn-north-4-hcs-cloud-cdn' (obs-image-desens), 'cn-south-1-cloud-res-model-sdk' (dds-dcs-instance-management), 'cn-north-4-hcli-cloud' (waf-aad-rule-management), and 'hcloud' (architecture-resource-archive). Because OBS bucket names share a global namespace across all Huawei Cloud customers, any unclaimed bucket name could be registered by an attacker to serve malicious install scripts.
Fix: Use only the verified official Huawei Cloud KooCLI download domain, pin release versions, download archives to disk first, and verify SHA-256 checksums before executing any installer.
-
LowTroubleshooting and setup guides disable TLS verification or edit /etc/hosts
skills/monitoring/ces/huawei-cloud-ces-aom-capacity-assessment/references/troubleshooting-dns.md:73
When regional CES or AOM endpoints time out, troubleshooting-dns.md instructs either appending public IPs to the system /etc/hosts file or calling hcloud against a raw IP with '--cli-skip-secure-verify=true' (and 'curl -sk'). Similar TLS-bypass flags appear in huawei-cloud-sac-intl-find-and-deploy ('--insecure' when downloading Terraform templates) and huawei-cloud-migration-windows-ecs ('verify_cert = false'). Disabling TLS certificate validation exposes authenticated cloud API calls or downloaded Terraform code to man-in-the-middle interception and tampering.
Fix: Remove instructions that disable TLS verification ('--cli-skip-secure-verify=true', 'curl -k', '--insecure', 'verify_cert = false') and fix VPC/DNS routing properly instead of mutating /etc/hosts.
-
LowUnencrypted WinRM Basic authentication advised for public-network test migrations
skills/migrationom/oos/huawei-cloud-migration-windows-ecs/references/sms-agent-guide.md:71
Both sms-agent-guide.md (with a comment permitting public-internet migration test environments) and troubleshooting.md provide PowerShell commands that enable WinRM Basic authentication alongside AllowUnencrypted='true' on TCP port 5985. Using Basic authentication without encryption over a public or shared network transmits the Windows Administrator password and subsequent commands in cleartext.
Fix: Remove the AllowUnencrypted='true' configuration snippet and require WinRM over HTTPS (port 5986) or an encrypted VPN/SSH tunnel even in test environments.
-
LowOpt-out telemetry downloads a binary at runtime and sends usage reports by default
skills/network/ops/huawei-cloud-sg-manage/references/cli-installation-guide.md:18
Several skills (including sg-manage, vpc-network-diagnosis-management, cts-trace-management, smn-dms-message, dew-key-management, iam-diagnose, iam-manage, waf-aad-rule-management, dds-dcs-instance-management, sac-intl-find-and-deploy, and obs-lifecycle-management) run ensure_cli.sh or install_cli.sh to download skill-quality-cli into ~/.local/bin/ and wrap commands so execution telemetry is sent automatically to skillsapi.developer.myhuaweicloud.com unless the user sets SKILL_QUALITY_REPORT=0 or SKILL_QUALITY_DISABLE=1.
Fix: Make telemetry strictly opt-in rather than opt-out, and prompt the user before downloading external telemetry binaries at runtime.
-
LowCredentials passed directly in CLI arguments expose secrets in process lists and history
skills/security/hss/huawei-cloud-hss-query/references/cli-installation-guide.md:25
Multiple guides instruct passing sensitive credentials directly on the command line—such as '--cli-access-key' and '--cli-secret-key' here (recommended for scripts/CI) as well as in network-topology-audit and oa-risk-inspection, '--password={password}' in dds-dcs-instance-management (including inside skill-quality-cli wrapper examples), '-p <login-secret>' in cloudrobo-dataset docker login, and 'hcloud obs config -i=... -k=...' in obs-dir-count. Command-line arguments are visible to other local users via the process table and are saved in shell history and CI logs.
Fix: Use interactive prompts ('hcloud configure init'), '--password-stdin' for Docker login, protected JSON input files ('chmod 600'), or environment variables instead of command-line secret flags.
-
LowOver-broad IAM policies recommended in several skill guides
skills/solution/sac/huawei-cloud-sac-yolo/references/iam-policies.md:117
Several IAM policy guides recommend excessive privileges beyond what the skill needs: huawei-cloud-sac-yolo instructs granting the 'Tenant Administrator' role across 'All resources' to an RFS agency and uses wildcard service actions ('ecs:servers:*', 'vpc:vpcs:*'); huawei-cloud-ces-elb-monitoring includes 'elb:*' and 'ces:*' in a monitoring policy; huawei-cloud-ecs-alert suggests 'ECS FullAccess' when troubleshooting 403 errors despite only reading ECS instances; and huawei-cloud-dds-dcs-instance-management offers 'dds:*:*' and 'dcs:*:*' on '*'.
Fix: Replace 'Tenant Administrator', '*FullAccess', and service-wide '*' wildcards with the minimum specific actions and scoped resources required for each skill.
-
LowAgent instructions template uploads repository contents without asking
This template for AGENTS.md instructs the AI agent to proactively call add_resource with the full repository path whenever starting work in a repository, explicitly telling it not to wait to be asked, and to use recalled memories across sessions and projects without mentioning the retrieval process. Installing this snippet can send proprietary source code, .env files, and local secrets in any opened repository to the OpenViking server without user confirmation, and allows prompt-injection text inside an untrusted repository to persist in shared memory across unrelated projects.
Fix: Require explicit user confirmation before indexing a repository with add_resource, exclude sensitive files (.env, keys, credentials), and scope recalled memories per project.
-
LowSecurity best-practices template grants public read and write access to OBS bucket
Under the 'Example: Secure OBS Bucket' section, the reference template for huaweicloud_obs_bucket_acl includes a public_permission block granting anonymous users both READ and WRITE access to the bucket. When an agent uses this security reference to generate bucket ACL code, anyone on the internet can read, overwrite, or upload objects in a bucket meant to be hardened.
Fix: Remove the public_permission block from the Secure OBS Bucket ACL example so anonymous public read and write access is not granted.
-
LowOMS migration template makes private source OBS bucket publicly readable
In both the task-group and single-task migration workflows, the template creates a private KMS-encrypted OBS bucket and then attaches a huaweicloud_obs_bucket_policy granting GetObject to Principal '*' on all objects in the bucket, even though the OMS migration resource already authenticates with access_key and secret_key. Generating migration Terraform from this template exposes private source bucket data to anyone on the internet.
Fix: Remove the wildcard Principal bucket policy or restrict Principal to the specific migration service/account principal instead of '*'.
-
LowAPI Gateway custom authorizer template accepts any non-empty token as admin
In the 'Deploy API with Custom Authentication' reference template, the inline FunctionGraph authorizer checks only whether the Authorization header is non-empty and then immediately returns statusCode 200 with userRole set to 'admin' after a placeholder comment. If generated and applied without replacing the stub, the public API Gateway endpoint accepts any arbitrary string as a valid admin token.
Fix: Make the template authorizer fail closed by returning statusCode 401 until real token validation logic is implemented.
-
LowMRS host diagnosis skills include user creation and deletion API definitions
skills/bigdata/mrs/huawei-cloud-mrs-host-alarm-diagnose/scripts/manager_api_apis/permission.yaml:23
Both huawei-cloud-mrs-host-alarm-diagnose and huawei-cloud-mrs-host-fault-diagnose define create_user (POST /api/v2/permission/users) and delete_user (DELETE /api/v2/permission/users) in their Manager API YAML catalogs despite being diagnostic skills. Exposing mutating user-management operations increases impact if an agent is tricked by prompt injection inside cluster logs or alarms.
Fix: Remove create_user and delete_user from permission.yaml in both diagnostic skills so only read-only inspection endpoints remain.
-
LowGlobal error hooks activate VOD collector on unrelated tool failures
skills/devtools/common/huawei-cloud-vod-collector/hooks/openclaw/hooks.yaml:2
Both hooks/openclaw/hooks.yaml and hooks/hermes/hooks.yaml register high-priority PostToolUse / on_tool_result hooks that activate huawei-cloud-vod-collector whenever any tool exits with a non-zero code and 'error' in stderr (or HTTP 4xx/5xx), with no filter restricting activation to Huawei Cloud tools or skills. This causes the skill to intercept errors from completely unrelated projects and commands.
Fix: Add a condition scoping the hook to Huawei Cloud CLI/SDK commands or skills rather than triggering on all failed tool calls.
-
LowTerraform Kafka plugin template strips sensitivity masking from passwords
The API Gateway Kafka forward plugin template wraps var.kafka_sasl_password and var.kafka_password in Terraform's nonsensitive() function, overriding sensitive = true on those variables. This causes plaintext SASL passwords to be printed in terminal output, CI logs, and agent tool results during terraform plan and terraform apply.
Fix: Remove nonsensitive() around var.kafka_sasl_password and var.kafka_password so Terraform keeps the plugin content redacted in plan and apply output.
-
LowCCE Terraform template writes kubeconfig without restrictive file permissions
The template writes huaweicloud_cce_cluster.test.kube_config_raw (which contains cluster client credentials and private keys) to .kube/config using local_file without setting file_permission. In hashicorp/local, local_file defaults file_permission to 0777 (masked only by umask), which can leave cluster admin credentials readable by other local users on shared machines.
Fix: Use local_sensitive_file or explicitly set file_permission = "0600" and directory_permission = "0700" when writing .kube/config.
-
LowMutating and container-exec APIs misclassified as read-only in service catalog
skills/solution/ops/huawei-cloud-solution-ops/references/hcloud-service-catalog/cci.json:1
The hcloud service catalog marks operations as read_only based on name keywords such as Get, List, and Check, which misclassifies several state-changing or command-execution APIs as read_only: true. Specifically, CCI connectCoreV1GetNamespacedPodExec (/api/v1/namespaces/{{namespace}}/pods/{{name}}/exec, which runs commands inside a container), BCS HandleUnionMemberQuitList in bcs.json (PUT /v2/{{project_id}}/members/quit, which removes a member from a blockchain consortium), and CloudIDE CheckMaliciousExtensionEvaluation in cloudide.json (POST /v1/marketplace/extension/evaluation/accusation) are all marked read_only: true. Any query tool or guardrail in huawei-cloud-solution-ops that trusts this flag can allow these operations without change-plan confirmation.
Fix: Derive read_only from the HTTP method and actual API behavior rather than name substrings alone, and set read_only to false for connectCoreV1GetNamespacedPodExec, HandleUnionMemberQuitList, and CheckMaliciousExtensionEvaluation.
-
LowTerraform security group template defaults ingress to 0.0.0.0/0 and opens SSH, Telnet, and RDP
In the Deploy Security Group reference used by the Terraform generator skill, both the security_group_rule_configurations variable and the huaweicloud_networking_secgroup_rule resource default direction to ingress, remote_ip_prefix to 0.0.0.0/0, and protocol/ports to null (all traffic). In addition, the accompanying terraform.tfvars preset explicitly opens TCP ports 22-23 (SSH and unencrypted Telnet) and 3389 (RDP) to 0.0.0.0/0, encouraging generated deployments to expose administrative services to the entire internet.
Fix: Remove the 0.0.0.0/0 default on ingress rules, require an explicit trusted CIDR block, and remove public 0.0.0.0/0 exposure of ports 22, 23, and 3389 from the example tfvars preset.
-
LowTerraform reference templates place plaintext cloud credentials in terraform.tfvars
Several Terraform best-practice templates in VPC-best-practices.md and WAF-best-practices.md instruct writing Huawei Cloud access_key and secret_key (as well as administrator_password and VPN pre-shared keys/passwords in VPN-best-practices.md) directly into terraform.tfvars in the working directory. Generating plaintext secret keys in project files increases the risk of committing long-lived cloud credentials to version control.
Fix: Configure Huawei Cloud authentication via environment variables (such as HW_ACCESS_KEY and HW_SECRET_KEY) or shared credential profiles, and mark password/PSK variables as sensitive = true without hardcoding them in tracked tfvars templates.
-
LowMutating cloud operations misclassified as read-only in service catalog
skills/solution/ops/huawei-cloud-solution-ops/references/hcloud-service-catalog/codeartsrepo.json:1
Several state-changing PUT and POST operations in the hcloud service catalog are marked read_only: true because their operation names contain words like List, Show, or Download. Notably, ListTemplatesTwo (PUT /v2/repositories/{{repository_uuid}}/template-status) sets whether a CodeArts repository is public or private, ShowHasPipeline (PUT /v1/repositories/{{repository_uuid}}/pipeline) modifies repository pipeline status, SaveBackupDownloadPolicy in dds.json (POST /v3/{{project_id}}/backups/download-policy) changes database backup download policies, and ListTaskParameter in codeartscheck.json and codecheck.json (POST /v2/{{project_id}}/tasks/{{task_id}}/config-parameters) updates check parameters. Any agent guardrail or workflow relying on read_only: true to skip confirmation can be tricked into modifying repository visibility or security settings.
Fix: Derive read_only in the catalog generator from both the HTTP method and actual API behavior rather than operation name prefixes alone, and set read_only to false for ListTemplatesTwo, ShowHasPipeline, SaveBackupDownloadPolicy, and ListTaskParameter.
-
LowMutating API operations marked as read_only in service catalog
skills/solution/ops/huawei-cloud-solution-ops/references/hcloud-service-catalog/modelarts.json:1
In the hcloud service catalog used by huawei-cloud-solution-ops, operations whose names begin with 'Show' are marked '"read_only": true' even when they perform state-changing POST actions. Specifically, 'ShowAutoSearchTrialEarlyStop' in modelarts.json sends a POST request that stops a running hyperparameter search trial, and 'ShowCustomConnector' in mssi.json sends a POST request ('/release') that publishes a custom connector. Any workflow or safety gate relying on the catalog's 'read_only' flag to permit unconfirmed execution could invoke these mutating calls without approval.
Fix: Set '"read_only": false' for 'ShowAutoSearchTrialEarlyStop' in modelarts.json and 'ShowCustomConnector' in mssi.json, and classify 'read_only' using HTTP method and endpoint behavior rather than name prefixes alone.
-
LowPersistent unencrypted Docker TCP 2375 listener paired with non-persistent iptables rules
skills/solution/ops/huawei-cloud-solution-ops/references/playbooks/docker-remote-api-readiness.md:54
When the user requests Docker Remote API access, this cloud-init template writes "tcp://${DOCKER_TCP_BIND}:2375" to /etc/docker/daemon.json while applying the source-CIDR restriction only via in-memory iptables commands, and instructs testing via http://<EIP>:2375. Contrary to the note claiming the listener is not persisted across reboots, /etc/docker/daemon.json persists on every boot while the unsaved iptables DROP rule disappears after a reboot, leaving an unauthenticated root-equivalent Docker API exposed to anyone permitted by the cloud security group.
Fix: Require mutual TLS (port 2376 with --tlsverify) instead of unauthenticated TCP 2375, and persist any host firewall rules across reboots using netfilter-persistent or nftables.
-
LowUnencrypted WinRM HTTP configured for public-internet Windows migration template
The public-internet Windows migration configuration template (scenario: public, connecting to a source server's public IP as Administrator) defaults to winrm_port: 5985 and winrm_use_ssl: false. Running WinRM without TLS over the public internet lacks server certificate verification and exposes administrative traffic to downgrade or man-in-the-middle risks.
Fix: Default public-internet migration templates to WinRM over HTTPS (port 5986, winrm_use_ssl: true) with certificate validation, or tunnel WinRM over an encrypted SSH/VPN channel.
-
LowKafka SASL password explicitly unmarked sensitive via nonsensitive()
The APIG Kafka forward plugin template wraps var.kafka_sasl_password and var.kafka_password in nonsensitive() when constructing the plugin JSON content. This strips Terraform's built-in secret redaction and causes plaintext Kafka credentials to be printed in terraform plan and apply output (including AI agent transcripts).
Fix: Remove nonsensitive() calls so Terraform keeps the entire rendered plugin content marked sensitive in plan and apply output.
-
LowCloud secret key and proxy password copied to Windows clipboard without clearing
While the script deletes its temporary base64 credential file to limit exposure, _InputRelative copies the Huawei Cloud AK, SK, and proxy password into the Windows clipboard via ClipPut($text) to paste them into the GUI and never clears the clipboard afterward. If the enterprise-project popup is skipped or RDP clipboard redirection or clipboard history is active, the secret key remains readable in the session clipboard.
Fix: Call ClipPut("") immediately after pasting and verifying each sensitive field, and again before exiting the script.
-
LowWSS inference test client unconditionally disables TLS verification
skills/ai/cloudrobo/huawei-cloud-cloudrobo-inference-adapter/scripts/test_client.py:183
InferenceTestClient hardcodes ssl.CERT_NONE and check_hostname = False for all wss:// connections (the same code also appears in huawei-cloud-cloudrobo-model-adapt-workflow/scripts/test_client.py). While harmless for 127.0.0.1, passing a remote --host exposes test images and prompts to man-in-the-middle interception with no flag to enable certificate verification.
Fix: Verify TLS certificates by default for non-loopback hosts and add an explicit --insecure or --ca-cert option for self-signed certificates.
-
LowOMS Terraform example attaches a public-read policy to a private encrypted bucket
skills/solution/ops/huawei-cloud-solution-ops/examples/terraform/oms_stack/main.tf:40
In the OMS migration Terraform template, bucket_acl defaults to private and KMS encryption is enabled, but main.tf unconditionally attaches an huaweicloud_obs_bucket_policy granting GetObject to Principal '*' on all objects in the source bucket. A user adapting this example template would unknowingly expose all objects in the bucket to anonymous public download.
Fix: Remove the wildcard Principal '*' bucket policy or restrict Principal to the specific IAM account or agency performing the migration.
-
LowFunctionGraph Terraform example uses vulnerable Pillow 7.1.2, unsanitized object paths, and over-broad agency
skills/solution/ops/huawei-cloud-solution-ops/examples/terraform/fgs_stack/variables.tf:123
The FunctionGraph image-compression Terraform example deploys Python 3.6 with pillow-7.1.2 (which has known code-execution vulnerabilities when parsing untrusted images), concatenates URL-decoded OBS object keys directly onto /tmp/<rand>/ without checking for '../' path traversal, and suggests function_all_trust in terraform.tfvars.example.
Fix: Sanitize src_object_key so resolved paths stay inside download_dir, upgrade the runtime and Pillow dependency to a supported patched version, and scope the example IAM agency to least-privilege OBS bucket access.
