← MCP server and skill grades

skills

google/skills/skills

D
Type
Skill
Scanned
Nov 13, 2026
Version
67c4dbc8ff63
Files read
1432

This collection gives AI coding assistants instructions, templates, and helper scripts to build, deploy, and troubleshoot workflows across Google Cloud, Firebase, Google Ads, Google Analytics, Genkit, and Android AI. Its biggest risks are setup commands that run unencrypted HTTP downloads directly in your terminal (allowing a network attacker to run code on your computer), helper scripts that can send live Google Cloud tokens to untrusted servers, and templates that deploy cloud services publicly without authentication.

What it can reach

Shell commands
Runs local terminal commands (including gcloud, kubectl, terraform, docker, adb, npm, npx, pip, git, and curl) and remote SSH commands on Google Cloud jump hosts and Colab VMs.
Files
Reads and writes project source code, infrastructure configs, AI models, reports, temporary files in /tmp, shell profile files, AI assistant MCP settings (such as Cursor, Claude Desktop, and VS Code), and remote Filestore NFS shares.
Network
Connects over HTTPS, HTTP, and WebSockets to Google Cloud, Firebase, Google Ads, Google Analytics, Hugging Face, GitHub, package registries (npm, PyPI, Maven), and custom MCP or model servers, and opens local ports for web servers and device forwarding.
Secrets
Reads and generates Google Cloud credentials and OAuth tokens, and handles OAuth client secrets, refresh tokens, Secret Manager values, and API keys (such as Gemini, OpenAI, Anthropic, and Hugging Face) in environment variables and config files.
System
Communicates with USB-attached or remote Android devices via ADB to run benchmarks and commands, and installs CLI tools and software packages on your machine.
Browser
Opens your web browser for Google Cloud and Firebase login flows and generates local HTML reports to view in a browser.

Findings

2 high6 medium20 low
  1. HighCleartext HTTP and unpinned remote install scripts piped to bash

    skills/cloud/developing-genkit-go/SKILL.md:139

    The Genkit CLI installation instructions in both developing-genkit-go and developing-genkit-dart run curl -sL cli.genkit.dev | bash without an https:// scheme. Because curl defaults to cleartext HTTP on port 80 when no scheme is given, anyone on the same network (such as public Wi-Fi) can intercept the initial request before any redirect and feed arbitrary commands directly into bash. Several other skills also run unpinned remote scripts or manifests without version or checksum verification, including bash <(curl -sSL https://storage.googleapis.com/cloud-samples-data/adc/setup_adc.sh) in skills/cloud/agent-platform-migrate-from-ai-studio/SKILL.md and kubectl apply -f against the mutable master branch in skills/cloud/gke-batch-hpc/SKILL.md.

    Fix: Always specify https:// explicitly, pin installer scripts and Kubernetes manifests to immutable version tags or commit digests, verify SHA-256 checksums before execution, or install pinned packages via a package manager after user confirmation.

  2. HighGCP OAuth access token sent to unvalidated hostname in helper scripts

    skills/cloud/agent-platform-eval-flywheel/scripts/endpoint_evaluation.py:103

    endpoint_evaluation.py builds its target HTTPS URL directly from the --dedicated_endpoint_dns argument (and --location when omitted) without verifying that the hostname belongs to Google Cloud, then attaches a live gcloud auth print-access-token Bearer token and posts every prompt from the evaluation dataset to that URL. Similarly, skills/cloud/agent-platform-skill-registry/scripts/skill_registry_ops.py interpolates --location / GCP_LOCATION directly into https://{region}-aiplatform.googleapis.com without validating the region format, allowing a crafted value such as attacker.example.com? to redirect requests carrying Application Default Credentials Bearer tokens even on read-only search or list actions. If prompt injection tricks the agent into passing an attacker-controlled host, the user's Google Cloud OAuth token is exfiltrated.

    Fix: Validate that dedicated_endpoint_dns ends with an expected Vertex AI domain suffix (such as .prediction.vertexai.goog or -aiplatform.googleapis.com) with no URL-breaking characters, and validate location/region arguments against ^[a-z0-9-]+$ before sending Authorization headers.

  3. MediumShared /tmp virtual environment reuse allows local code execution

    skills/cloud/cloud-logging-finops/SKILL.md:101

    The Cloud Logging FinOps skill creates a virtual environment at the fixed, world-writable path /tmp/finops_venv only if /tmp/finops_venv/bin/python does not already exist, and then runs scripts with /tmp/finops_venv/bin/python using the user's Google Cloud credentials. On a shared multi-user machine, another local user can pre-create /tmp/finops_venv/bin/python with malicious code before the victim runs the skill, causing the check to skip environment creation and execute the attacker's binary as the victim. Other skills such as skills/cloud/iam-helper-for-policy-simulator/SKILL.md similarly read and write sensitive IAM policy JSON files at fixed paths in /tmp without per-user isolation.

    Fix: Create reusable virtual environments inside a user-owned directory such as ~/.cache/finops_venv (mode 0700), and use mktemp -d for temporary IAM policy and simulation files instead of fixed /tmp paths.

  4. MediumGoogle Cloud identity token sent to unvalidated bridge URL

    skills/cloud/google-cloud-filestore-nfs-browser/scripts/nfs_browser.py:65

    HTTPBridgeEngine accepts any URL via --bridge-url (or FILESTORE_BRIDGE_URL) without checking that it uses HTTPS or belongs to an expected Cloud Run domain, and automatically attaches a live bearer token from FILESTORE_BRIDGE_TOKEN or gcloud auth print-identity-token. If an agent is tricked by prompt injection into passing an attacker-controlled or plaintext HTTP URL to --bridge-url, the user's Google Cloud identity token is sent directly to that server.

    Fix: Require https:// on bridge_url, restrict allowed hostnames (such as *.run.app or an explicit allowlist), and pass --audiences=<bridge_url> to gcloud auth print-identity-token so tokens are audience-bound.

  5. MediumSymlink escape during content search in Filestore NFS bridge server

    skills/cloud/google-cloud-filestore-nfs-browser/scripts/bridge_server/main.py:165

    While safe_path() blocks direct reads of symlinks that resolve outside MOUNT_ROOT, search_files() only validates the starting directory and then opens every file yielded by os.walk() without checking os.path.realpath() on each file. Because os.walk() includes symlinked files in its file list, any untrusted user or workload with write access to the shared NFS volume can plant a symlink to sensitive container or host files (such as /proc/self/environ or credentials) and have their contents returned when a grep search runs. The same flaw also exists in jump_host_engine.py's remote search script.

    Fix: Skip symbolic links with os.path.islink(fp) or validate each walked file path with safe_path() before calling os.path.getsize() or open().

  6. MediumPath traversal escapes NFS mount on remote GCE jump host

    skills/cloud/google-cloud-filestore-nfs-browser/scripts/jump_host_engine.py:51

    The remote read and stat helpers check os.path.realpath() and os.path.commonpath() to keep access inside the NFS mount directory, but _build_remote_tree_script() and _build_remote_search_script() omit that check completely. If an agent is tricked into passing a path like /../../etc or ../../home to the tree or search subcommands, the script lists directories and searches file contents across the entire GCE jump host VM filesystem.

    Fix: Enforce the same os.path.realpath() and os.path.commonpath([mount_base, root]) == mount_base check on root and on every walked entry in _build_remote_tree_script() and _build_remote_search_script().

  7. MediumPlaintext HTTP install script piped directly to bash

    skills/cloud/developing-genkit-go/references/getting-started.md:104

    The Genkit CLI installation instruction runs curl against cli.genkit.dev without specifying https:// and pipes the response directly into bash. Because curl defaults to plaintext http:// when no scheme is given, an attacker on the local network or transit path can intercept the initial request before any HTTPS redirect and return arbitrary shell commands that execute immediately on the developer's machine. Even over HTTPS, piping an unpinned script into a shell runs unverified remote code.

    Fix: Install a pinned version via npm (for example, npm install -g genkit-cli@1.29.0) or enforce HTTPS with curl --proto '=https' --tlsv1.2 -sSfL https://cli.genkit.dev and verify a SHA-256 checksum before executing.

  8. MediumUnescaped telemetry fields rendered via innerHTML in migration report

    skills/developers/migrate-to-rust/references/infographic_template.html:451

    When rendering migration_infographic.html, the script interpolates telemetry strings—including file paths, opportunity titles and descriptions, test failure summaries, and parity notes collected while auditing a legacy codebase—directly into DOM elements via innerHTML without HTML escaping. If a user migrates an untrusted repository containing HTML or script payloads in filenames, comments, or test output, opening the generated HTML report in a browser executes arbitrary JavaScript.

    Fix: HTML-escape &, <, >, ", and ' in all dynamic values before inserting them into innerHTML, or populate DOM nodes using textContent.

  9. LowPath traversal via eval_case_id when rendering trajectories to disk

    skills/cloud/agent-platform-autorater-customization/scripts/render_agent_data.py:90

    When --out_dir is passed, render_agent_data.py reads eval_case_id from each JSONL record and passes it directly into os.path.join(a.out_dir, f"{cid}.txt") before opening the file for writing. If an untrusted trajectory JSONL file contains ../ segments or a leading / in eval_case_id (which causes Python's os.path.join to discard a.out_dir), running the script overwrites .txt files outside the intended output directory.

    Fix: Sanitize eval_case_id to safe filename characters (or apply os.path.basename) and verify with os.path.realpath that the resolved destination remains inside os.path.realpath(a.out_dir) before writing.

  10. LowUnconfirmed organization- and billing-wide admin role self-escalation on permission errors

    skills/cloud/google-cloud-recipe-foundation-builder/SKILL.md:181

    While Phase 1 asks the user to approve creating folders, projects, org policies, and logging sinks, Phase 2 instructs the agent to automatically handle any Permission Denied error by granting entire administrative role bundles (up to 9 Organization Admin roles, 9 Security Admin roles, and 3 Billing Admin roles) to the active identity at the Organization and Billing Account root without pausing for user confirmation or removing the roles afterward. This silently and permanently expands privileges far beyond least privilege.

    Fix: Do not autonomously grant Organization, Security, or Billing Admin role groups on failure; halt, report the minimal required IAM role, and require explicit user confirmation before running any add-iam-policy-binding command.

  11. LowSkill instructs exporting default Compute Engine service account keys and disabling org key-creation guardrails

    skills/cloud/agent-platform-migrate-from-ai-studio/SKILL.md:289

    For OpenClaw deployments, this skill instructs creating a downloadable private key file for the Compute Engine default service account (which commonly holds broad Editor permissions), overwriting ~/.config/gcloud/application_default_credentials.json, copying the key to a VM, and prompting the user to disable Google Cloud's iam.disableServiceAccountKeyCreation Organization Policy (mislabeling the security guardrail as a legacy policy). Long-lived exported keys for the default Compute Engine service account create a high credential-theft risk and are unnecessary on GCE VMs, which authenticate natively through attached service accounts.

    Fix: Attach a dedicated, least-privilege service account to the Compute Engine VM and use the instance metadata server (or service account impersonation locally) instead of creating service account key files or disabling iam.disableServiceAccountKeyCreation.

  12. LowSilent modification of global agent MCP configuration without user approval

    skills/cloud/cloud-monitoring-metric-selection/SKILL.md:65

    When a user asks to find or list Cloud Monitoring metrics and the list_metric_descriptors MCP tool is missing, Step 1 instructs the agent to locate and directly edit global IDE or agent MCP configuration files (such as ~/.gemini/config/mcp_config.json or claude_desktop_config.json) without asking for permission first, registering a remote MCP server with google_credentials authentication. Modifying global agent configuration files crosses a trust boundary and should require explicit user consent.

    Fix: Prompt the user for explicit confirmation before modifying any IDE or agent MCP configuration file, or print the configuration snippet for the user to add manually.

  13. LowPython code injection via unescaped CLI arguments in Apigee governance scripts

    skills/cloud/apigee-agent-gateway-governance/scripts/deploy_proxy.sh:72

    deploy_proxy.sh and ensure_apigee_environment.sh interpolate command-line arguments such as $ENV, $PROXY, $ORG, and $REGION directly into double-quoted python3 -c code strings instead of passing them through sys.argv. Even in default dry-run mode (without --yes), a prompt-injected argument containing a single quote breaks out of the Python string literal and executes arbitrary code.

    Fix: Pass shell variables as positional arguments to python3 -c and read them via sys.argv[1:] instead of interpolating them into Python source strings.

  14. LowUnescaped policy fields enable shell injection in generated gcloud commands

    skills/cloud/cloud-armor-crs-rule-upgrader/scripts/upgrade_policy.py:1117

    When --format=gcloud is used, generate_gcloud_commands() wraps policy_name, upgraded_expression, action, and rule description in double quotes without escaping double quotes, backticks, or $() substitutions. If a user or agent upgrades a policy JSON from an untrusted or shared source and runs the printed gcloud commands, shell metacharacters inside a rule description or CEL expression will execute as shell commands.

    Fix: Use shlex.quote() on pol_name, expr, action, and desc when constructing the gcloud CLI command strings.

  15. LowFilestore troubleshooting remediation defaults to NO_ROOT_SQUASH and 10.0.0.0/8

    skills/cloud/google-cloud-filestore-log-troubleshooting/scripts/diagnose_lib.py:197

    When building or applying fixes (--apply-fix) for Filestore mount failures, build_non_destructive_export_json() hardcodes squash-mode to NO_ROOT_SQUASH with READ_WRITE access, giving remote root users full superuser privileges on the NFS share. In addition, if --client-ip and --client-subnet are omitted, firewall remediation defaults to opening NFS ports 2049 and 111 to the entire 10.0.0.0/8 private address space.

    Fix: Default squash-mode to ROOT_SQUASH and require an explicit --client-ip or --client-subnet before generating or applying firewall and export ACL rules.

  16. LowRuntime download and device execution of unpinned binaries in LiteRT benchmark script

    skills/platform-devices/litert-google-tensor/scripts/benchmark_on_device.py:609

    benchmark_on_device.py downloads native ARM64 binaries (benchmark_model and libLiteRtDispatch_GoogleTensor.so) at runtime, pushes them to an attached Android phone over adb, marks them executable, and runs them without checking a pinned cryptographic SHA-256 hash. It also accepts arbitrary URLs via --benchmark-url and --dispatch-url without host or HTTPS validation.

    Fix: Verify downloaded binaries against pinned SHA-256 digests in assets/versions.json before pushing them to the device, and restrict download URLs to https://storage.googleapis.com/litert/binaries/ unless explicitly overridden by the user.

  17. LowDeployment guide instructs granting broad roles/aiplatform.admin as automatic fallback

    skills/cloud/agent-platform-deploy/references/copy_deploy_guide.md:110

    In Step 0.3.2, if adding the Vertex AI Service Agent role to the source project fails, the guide instructs the agent to grant roles/aiplatform.admin to the user on the destination project without listing IAM policy changes under the Tier M confirmation gate. roles/aiplatform.admin grants full administrative control over all Vertex AI resources in the project.

    Fix: Require explicit user confirmation before any gcloud projects add-iam-policy-binding command and recommend the least-privilege role needed for the operation.

  18. LowLocal ADB forwarder accepts unauthenticated localhost connections without packet validation

    skills/cloud/developer-device-platform-basics/scripts/demo_adb_forwarder.py:103

    AdbForwarder listens on 127.0.0.1:<ephemeral_port> for 15 minutes and bridges raw ADB OPEN and WRTE streams directly to the user's authenticated Google Cloud device session without ADB AUTH verification, magic header checks, or disconnecting on unrecognized non-ADB traffic. On a shared machine, any local user or process can connect to the port and run shell commands on the remote Android device.

    Fix: Validate the ADB header magic field (magic == cmd ^ 0xFFFFFFFF), close the socket immediately on unrecognized commands such as HTTP requests, and restrict access on multi-user systems.

  19. LowPredictable /tmp file path used for IAM simulation results

    skills/cloud/iam-helper-for-policy-simulator/scripts/analyze_simulation.py:10

    analyze_simulation.py reads IAM Policy Simulator output from a hardcoded /tmp/simulation_results.json path instead of a user-owned or mktemp file. On a shared multi-user machine, another local user could pre-create or tamper with that file to spoof simulation results such as hiding access revocations.

    Fix: Accept the simulation results path as a CLI argument created with mktemp or read JSON from standard input.

  20. LowUnpinned third-party npm package executed automatically with npx -y

    skills/cloud/bigquery-migration-sql-translator/references/mcp_config.md:89

    Option A for installing the BigQuery Migration Service MCP configuration runs npx -y add-mcp@latest, which downloads and executes whatever code is currently published at @latest for a third-party npm package (neondatabase/add-mcp) without interactive confirmation. Because this tool runs locally and modifies installed AI coding agent configurations (~/.claude.json, ~/.cursor/mcp.json, ~/.gemini/settings.json), an upstream compromise of that npm package would execute arbitrary code on the developer's machine.

    Fix: Configure the MCP server JSON directly (Option B) or pin add-mcp to a vetted version without -y so the user confirms package installation.

  21. LowCloud Run Terraform template grants unauthenticated public access

    skills/cloud/cloud-run-basics/references/iac-usage.md:30

    The sole Terraform example for a Cloud Run service sets ingress = "INGRESS_TRAFFIC_ALL", disables deletion_protection, and binds roles/run.invoker to allUsers with no warning comment in the file. If an AI agent copies this template when scaffolding Terraform for a user's service, the deployed Cloud Run endpoint will be publicly invokable by anyone on the internet.

    Fix: Remove the allUsers IAM binding from the default template (or comment it out with an explicit warning) and bind roles/run.invoker only to specific authenticated principals.

  22. LowHardcoded weak initial password in AlloyDB Terraform template

    skills/cloud/alloydb-basics/references/iac-usage.md:45

    The Terraform example for creating an AlloyDB cluster hardcodes password = "alloydb-cluster" in plaintext inside initial_user and sets deletion_protection = false. If an agent reuses this pattern when generating infrastructure code, the database is created with a trivial password matching the cluster name and the credential is saved in version control and Terraform state.

    Fix: Use a sensitive Terraform input variable or a Google Cloud Secret Manager reference for the initial password instead of a hardcoded literal, and prefer IAM database authentication.

  23. LowBearer access token placed directly in project-level MCP config files

    skills/cloud/bigquery-migration-sql-translator/references/mcp_config.md:189

    Option B instructs putting a Google Cloud OAuth access token directly into project-level .mcp.json or .vscode/mcp.json files without warning the user to exclude those files from version control or interpolate environment variables. Project-level MCP config files are frequently committed to git repositories, which could expose a live bearer token before it expires.

    Fix: Use environment variable expansion in headers where supported, store credentials in user-scoped configuration files outside the repository, and remind users to add .mcp.json to .gitignore if tokens are present.

  24. LowIAP authorization extension template defaults to failOpen: true

    skills/cloud/google-cloud-solution-multi-agent-security/assets/agw-authz-extension.yaml:10

    The AuthzExtension manifest template for Identity-Aware Proxy (IAP) authorization sets failOpen: true. If the authorization service is unreachable, overloaded, or exceeds the 15-second timeout, Agent Gateway will allow requests through without an authorization check instead of blocking them. Other troubleshooting documentation in this repository explicitly flags failOpen: true as a security risk.

    Fix: Set failOpen: false so that Agent Gateway fails closed and denies requests whenever the authorization service errors or times out.

  25. LowNonce replay check bypassed when auth_nonce cookie is omitted

    skills/identity/sign-in-with-google-web/references/bff_fastapi_verification.md:78

    In the recommended FastAPI Backend-for-Frontend (BFF) authentication endpoint, auth_nonce_cookie is optional (defaulting to None) and the replay check only runs when auth_nonce_cookie is truthy. An attacker replaying a captured ID token or performing Login CSRF against /api/auth/google can simply omit the auth_nonce cookie from the POST request to skip nonce verification completely and obtain a valid session cookie.

    Fix: Reject the request whenever auth_nonce_cookie or the token's nonce claim is missing, compare them using secrets.compare_digest, and ensure the frontend passes the server-issued nonce to google.accounts.id.initialize.

  26. LowUnauthenticated agent HTTP endpoints exposed with wildcard CORS

    skills/cloud/developing-genkit-js/references/agents-deployment.md:86

    The Genkit HTTP deployment guide exposes agent execution, session snapshot (/getSnapshot), and abort routes without authentication on all network interfaces (app.listen(8080)) and configures Access-Control-Allow-Origin: *. Sibling Genkit serving guides for Dart, Go, and Python use similar 0.0.0.0 bindings and wildcard CORS defaults. If run on a developer workstation or server, any website visited in the user's browser or any peer on the local network can invoke the agent's tools, read session history, and consume paid model quota.

    Fix: Bind local development servers to 127.0.0.1, require authentication middleware or a contextProvider on all agent and snapshot routes, and restrict Access-Control-Allow-Origin to trusted application origins.

  27. LowUnpinned npx package execution at runtime without version or checksum verification

    skills/cloud/firebase-basics/references/local-env-setup.md:64

    Multiple skills instruct the agent to download and execute unpinned npm packages automatically via npx -y (such as firebase-tools@latest here, @philschmid/weather-mcp in gemini-api, and @modelcontextprotocol/server-sse-client with an OAuth bearer token in iam-helper-for-troubleshooting) or download binaries without checksum verification (mcp-toolbox in google-cloud-storage-basics). Automatically executing floating @latest or unpinned packages exposes the user to upstream supply-chain compromise.

    Fix: Pin exact package versions (as done for @toolbox-sdk/server@1.13.1) and verify cryptographic SHA-256 checksums for downloaded standalone binaries before marking them executable.

  28. LowMCP and skill setup guides download and execute unpinned packages at runtime

    skills/cloud/firebase-basics/references/setup/cursor.md:62

    The Firebase MCP setup guides (for Cursor, Antigravity, GitHub Copilot, Gemini CLI, and other agents) configure the MCP server to run 'npx -y firebase-tools@latest mcp', which silently downloads and executes whatever version is currently tagged @latest on npm every time the editor starts. Similarly, the Firebase setup and refresh guides run 'npx -y skills add firebase/agent-skills' without pinning the skills CLI version or repository commit, and the Google Ads MCP setup configures 'pipx run google-ads-mcp' without a version pin. If an upstream package is compromised, malicious code runs automatically on startup.

    Fix: Install vetted, version-pinned releases of firebase-tools, skills, and google-ads-mcp rather than invoking npx -y with @latest or unpinned pipx run on every client startup.