skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- 968f7bfaa016
- Files read
- 980
This project is a collection of Microsoft guides and helper scripts that help AI coding assistants build, deploy, and manage cloud apps and AI services on Azure. The biggest risk is that several included scripts and templates use unsafe settings that can open login-token servers to your local network, send live Azure tokens to unchecked web addresses, print secret keys into logs, or let AI tools act on Microsoft 365 accounts without asking for your approval.
What it can reach
- Files
- Reads and writes project source code, logs, environment (.env) files, cloud deployment templates, GitHub Actions workflows, and local state folders.
- Network
- Connects to Azure cloud services, Microsoft Graph, Microsoft Foundry, Azure OpenAI, GitHub APIs, external IP lookup services, and deployed application URLs.
- Shell commands
- Runs local terminal commands and bundled Bash, PowerShell, Python, and Node.js scripts using tools like Azure CLI, Kubernetes, Docker, Terraform, GitHub CLI, and curl.
- Secrets
- Reads local Azure login tokens and SSH public keys, accesses Azure Key Vault secrets, and creates or saves API keys, deployment tokens, and app passwords.
- Browser
- Opens Azure Portal deployment monitoring pages in your default web browser.
- System
- Installs global command-line tools, Azure CLI extensions, and Copilot plugins, and can launch privileged Kubernetes debugging containers when approved.
Findings
-
MediumEntra Agent ID sidecar Docker Compose binds unauthenticated token endpoints to 0.0.0.0
skills/entra-agent-id/references/sdk-sidecar-deployment.md:105
The Docker Compose template for the Microsoft Entra SDK for AgentID sidecar configures the sidecar with the Blueprint client secret and publishes port 5001:5000 without a loopback host address, binding 0.0.0.0:5001 on all host network interfaces. Because the sidecar exposes unauthenticated endpoints (/AuthorizationHeaderUnauthenticated/{name} and /DownstreamApiUnauthenticated/{name}), anyone on the same local network can request live Microsoft Graph OAuth 2.0 Bearer tokens or proxy downstream API calls with no authentication, contradicting the file's own warning to bind only to localhost.
Fix: Remove the host ports mapping from the sidecar service so only containers on the internal Docker network can reach http://sidecar:5000, or bind strictly to loopback using 127.0.0.1:5001:5000.
-
MediumFoundry helper scripts send live Entra ID bearer tokens to unvalidated endpoint hosts
skills/microsoft-foundry/foundry-agent/create/scripts/smoke_test_voice.py:220
smoke_test_voice.py (along with sibling scripts smoke_test_gpt_live.py and skills/microsoft-foundry/finetuning/scripts/common.py) acquires a live Entra ID access token via Azure CLI or DefaultAzureCredential and attaches it in an Authorization: Bearer header to whatever URL is supplied via --endpoint, --base-url, environment variables, or workspace .foundry/agent-metadata.yaml, without verifying that the hostname is an Azure domain (and common.py even allows plain http:// URLs). If prompt injection or an untrusted workspace metadata file points the endpoint to an external server, the script exfiltrates the user's Azure Entra token to the attacker.
Fix: Parse the endpoint URL, enforce https:// or wss://, and verify that the hostname ends with a trusted Azure suffix (such as .services.ai.azure.com, .openai.azure.com, or .cognitiveservices.azure.com) before attaching Entra ID bearer tokens.
-
MediumSilent plaintext HTTP lookup to ifconfig.me can let network attackers spoof VM firewall rules
skills/azure-compute/workflows/vm-creator/references/depth-probe/beginner.md:16
The VM Creator fast-path silently runs curl -s ifconfig.me without asking the user and without an https:// scheme, defaulting curl to unencrypted HTTP. Anyone on the local network (such as shared Wi-Fi) can intercept the HTTP response to whitelist their own IP address in the VM's SSH/RDP firewall rule, or block the request to trigger the fallback to * (0.0.0.0/0).
Fix: Use an explicit HTTPS endpoint with TLS verification (such as curl -fsS https://ifconfig.me), ask before contacting an external IP-echo service, and fail closed instead of falling back to *.
-
MediumNode.js Entra JWT middleware skips audience verification if environment variables are unset
skills/azure-prepare/references/services/app-service/templates/recipes/auth/source/nodejs.md:26
The Express auth recipe sets AUDIENCE from process.env.AZURE_APP_ID_URI || process.env.AZURE_CLIENT_ID without checking that it is defined. In jsonwebtoken, passing undefined or an empty string as audience disables aud claim validation while still trusting the shared tenant-wide issuer URL, allowing any valid Entra token for any other app in the same tenant to authenticate to the API.
Fix: Throw an error at module startup if AUDIENCE or process.env.AZURE_TENANT_ID is missing or empty so misconfigured deployments fail closed.
-
MediumWork IQ and MCP templates disable human approval on user-impersonating Microsoft 365 tools
Work IQ connects a Foundry agent to the signed-in user's Microsoft 365 mail, chats, meetings, and files via user-entra-token, exposing read and write/delete tools (do_action, create_entity, update_entity, delete_entity). Unlike the general MCP reference that requires require_approval: always, this template (along with tool-fabric-iq.md and the toolbox MCP templates) sets require_approval: never with no allowed_tools filter, allowing prompt injection from untrusted emails or web results to exfiltrate M365 data or run actions without confirmation.
Fix: Default to require_approval: always and encourage scoping allowed_tools to read-only operations unless the user explicitly opts out.
-
MediumAzure AI Search toolbox guide uses full admin key and passes it on the command line
skills/microsoft-foundry/foundry-agent/toolbox/references/tool-azure-ai-search.md:66
The Azure AI Search setup fetches the service's primary admin key (az search admin-key show) and passes it via --key on the azd ai connection create command line. This grants the runtime search tool full administrative rights to modify or delete indexes (contradicting the project's own AUTH-001 Entra rule) and exposes the secret in local process arguments.
Fix: Configure keyless Microsoft Entra ID role-based access for Azure AI Search connections by default, or use a read-only query key if key auth is explicitly required.
-
LowPython code injection via shell-interpolated arguments in discover_and_rank.sh
skills/microsoft-foundry/models/deploy-model/capacity/scripts/discover_and_rank.sh:60
The capacity sub-skill advertises discover_and_rank.sh as a read-only capacity check and instructs the agent to extract the model name, version, and minimum capacity from the user prompt and pass them as positional arguments. Inside discover_and_rank.sh, MIN_CAPACITY, MODEL_NAME, MODEL_VERSION, and JSON responses are interpolated directly into a double-quoted python3 -c code string instead of being passed via sys.argv or environment variables, allowing crafted arguments from prompt injection to execute arbitrary Python and OS commands.
Fix: Use a single-quoted python3 -c script and pass MODEL_NAME, MODEL_VERSION, MIN_CAPACITY, and JSON inputs via sys.argv, stdin, or environment variables rather than interpolating them into Python source code.
-
LowMandatory automatic installation of unpinned azd extensions when loading microsoft-foundry
skills/microsoft-foundry/scripts/check-and-setup-dependencies.sh:24
skills/microsoft-foundry/SKILL.md mandates running check-and-setup-dependencies.sh (or .ps1) immediately as the first step whenever the skill loads, before reading any sub-skill or workflow reference. The script automatically downloads and installs the microsoft.foundry and azure.ai.evaluations azd extensions without asking for user consent and without pinning versions, even when the user requested a strictly read-only or offline sub-skill such as validate or insights.
Fix: Do not auto-install extensions before checking whether the requested workflow needs them, prompt the user for confirmation before installing missing CLI extensions, and pin known-good minimum/exact versions where supported.
-
LowPlaintext storage and CLI/audit-log exposure of generated database passwords in azure-app-onboard
skills/azure-app-onboard/deploy/references/deploy-safety.md:38
Although blocked-patterns.md forbids writing secrets to disk or passing secrets inline on CLI arguments, deploy-safety.md instructs the agent to persist generated production passwords in plaintext to .copilot-azure/sessions/{id}/deploy-secrets.env without restrictive 0600 file permissions (retained for 7 days). In addition, code-deployment-container-apps.md and database-post-deploy.md pass database admin passwords directly on command lines (--parameters administratorLoginPassword=... and -p {pass}), while deploy-checklist-template.md logs every az command to deploy-audit.log and even suggests reading @secure() values back from deploy-audit.log.
Fix: Avoid writing plaintext secrets to workspace session files or audit logs; read secrets back from Azure Key Vault when needed, pass deployment secrets via stdin or protected 0600 parameter files deleted immediately after use, and redact secret arguments before writing deploy-audit.log.
-
LowShell command injection via eval of azd environment values in SQL post-deployment guides
skills/azure-deploy/references/recipes/azd/sql-managed-identity.md:16
sql-managed-identity.md (as well as post-deployment.md and verify.md) instructs the agent to run and generate postprovision hook scripts containing eval $(azd env get-values). Evaluating raw azd environment output in the shell performs word splitting, glob expansion, and command substitution on every environment variable and Bicep output, allowing shell metacharacters in untrusted templates or environment values to execute arbitrary commands.
Fix: Replace eval $(azd env get-values) with the line-by-line IFS= read -r parser already used in apply-migrations.sh and grant-sql-access.sh.
-
LowUnsanitized T-SQL identifier and literal interpolation in grant-and-migrate scripts
skills/azure-deploy/references/recipes/azd/scripts/grant-and-migrate.sh:69
grant-and-migrate.sh and grant-and-migrate.ps1 interpolate APP_NAME (loaded from SERVICE_WEB_NAME or SERVICE_API_NAME in azd env) directly into single-quoted T-SQL strings and bracketed identifiers executed with Entra SQL administrator privileges. Unlike skills/azure-prepare/references/services/sql-database/scripts/grant-sql-access.sh, which validates APP_NAME with a strict alphanumeric-hyphen regex and escapes quotes and brackets, these scripts perform no validation or escaping, enabling SQL injection if environment values are untrusted.
Fix: Validate APP_NAME against ^[a-zA-Z0-9][a-zA-Z0-9-]{0,58}[a-zA-Z0-9]$ and escape single quotes and closing brackets before constructing the T-SQL query, matching grant-sql-access.sh.
-
LowHardcoded fallback session secret in production Express configuration template
skills/azure-prepare/references/runtimes/nodejs.md:119
Under Complete Production Configuration, which generate.md instructs the agent to apply for containerized Node.js/Express deployments, express-session is configured with secret: process.env.SESSION_SECRET || 'dev-secret-change-in-prod'. If SESSION_SECRET is unset or misnamed in production app settings, the application silently starts and signs session cookies with a publicly known static string, allowing remote network attackers to forge session cookies against the deployed app.
Fix: Require SESSION_SECRET to be non-empty when NODE_ENV is production (throwing an error at startup if missing) instead of falling back to a hardcoded string.
-
LowShell eval of command strings in retry-az-create wrapper scripts
skills/python-appservice-deploy/scripts/retry-az-create.sh:36
retry-az-create.sh passes caller-provided command strings to eval (and retry-az-create.ps1 passes them to powershell -Command), while create-app.md instructs using user-supplied resource group, plan, and region names verbatim. If prompt injection supplies a resource name containing shell metacharacters, eval executes injected commands.
Fix: Validate resource group, plan, app, and region identifiers against strict Azure naming regexes and execute az commands directly as argument arrays without eval or powershell -Command.
-
LowUnpinned remote installation scripts, Kubernetes manifests, and build plugins
skills/azure-deploy/references/sdk/azd-deployment.md:8
Several skill references instruct running unpinned remote code or cluster definitions without version pinning or checksum verification, including curl -fsSL https://aka.ms/install-azd.sh | bash in azd-deployment.md, kubectl apply -f https://github.com/kubernetes-sigs/gateway-api/releases/latest/download/standard-install.yaml in skills/azure-kubernetes/azure-kubernetes-app-deploy/phases/quick-deploy.md, and OpenRewrite plugin version latest.release in skills/azure-upgrade/references/languages/java/scripts/upgrade_bom.py.
Fix: Prefer package managers (winget, brew, apt) or version-pinned release URLs with SHA-256 verification for CLI installers and Kubernetes CRDs, and pin explicit stable OpenRewrite plugin versions.
-
Lowaz CLI VM creation template leaves SSH or RDP open to the entire internet
skills/azure-compute/workflows/vm-creator/references/output-adapters/az-cli.md:25
While the VM Creator Plan Card and Bicep/Terraform templates restrict inbound SSH (port 22) or RDP (port 3389) to the user's IP address, the az CLI adapter passes a new NSG name directly to az vm create without restricting source addresses. Azure CLI then auto-creates a default allow rule open to * (0.0.0.0/0), exposing SSH or password-based Windows RDP to the entire internet.
Fix: Create the NSG and a source-IP-restricted rule explicitly with az network nsg create and az network nsg rule create before calling az vm create, or pass --nsg-rule NONE and add a scoped rule.
-
LowLocal agent startup example binds unauthenticated server to 0.0.0.0 on the host
skills/microsoft-foundry/foundry-agent/create/references/local-run.md:112
The local-run guide runs a hosted agent directly on the developer's computer using their local DefaultAzureCredential (az login or VS Code identity), but its startupCommand example binds uvicorn to 0.0.0.0 instead of 127.0.0.1. Because the local agent HTTP endpoint has no authentication, anyone on the same local network can send prompts to port 4001 and trigger the agent's Azure-authenticated tools.
Fix: Change the local startupCommand example to bind to 127.0.0.1 (--host 127.0.0.1) and warn against using 0.0.0.0 outside an isolated container.
-
LowFoundry resource creation workflow dumps raw API keys to terminal and LLM context
skills/microsoft-foundry/resource/create/references/workflows.md:168
Step 5 of the Foundry resource creation workflow (and Pattern A in patterns.md) instructs the agent to run az cognitiveservices account keys list immediately after creating an account. Executing this in an AI coding agent prints full-privilege key1 and key2 secrets to stdout and sends them into the AI model's conversation history without being asked.
Fix: Remove automatic key listing from the creation workflow, prefer Entra ID authentication, and instruct the user to retrieve keys directly only if key-based auth is explicitly needed.
-
LowKey Vault SDK reference examples log plaintext secret values to stdout
skills/azure-prepare/references/services/key-vault/sdk.md:17
The Node.js, Python, and .NET Key Vault SDK snippets print retrieved secret values (such as database connection strings) directly to stdout. Agents copying this reference into application code risk leaking Key Vault secrets into container logs and Application Insights.
Fix: Replace console.log(secret.value), print(secret.value), and Console.WriteLine(secret.Value) with comments showing safe in-memory consumption without logging the secret.
-
LowSpring Boot AKS guide exposes full Actuator health details publicly
skills/azure-kubernetes/azure-kubernetes-app-deploy/knowledge-packs/frameworks/spring-boot.md:39
Under Required configuration, the Spring Boot AKS pack sets management.endpoint.health.show-details=always while the Ingress template routes / publicly to the app. Unauthenticated callers can query /actuator/health to view internal component versions, database/cache status, and error details that Kubernetes probes do not need.
Fix: Remove management.endpoint.health.show-details=always or set show-details=when-authorized; probes only require management.endpoint.health.probes.enabled=true.
-
LowSQL Server Bicep template enables AllowAzureServices (0.0.0.0) firewall rule by default
skills/azure-prepare/references/services/sql-database/bicep.md:49
The base SQL Server Bicep template creates the AllowAzureServices rule (0.0.0.0 to 0.0.0.0), permitting network-level connections from any Azure subscription or tenant rather than only the user's resources (durable-task-scheduler/bicep.md similarly defaults ipAllowlist to 0.0.0.0/0). Although Entra ID authentication is still enforced, this broadens network reachability without warning.
Fix: Document that AllowAzureServices permits network traffic from other Azure tenants and recommend Private Endpoints or explicit VNet/IP rules for production.
-
LowPreset model deployment script writes to predictable /tmp/capacity_check.json path
skills/microsoft-foundry/models/deploy-model/preset/references/preset-workflow.md:200
Phase 5 writes region capacity output to a hardcoded /tmp/capacity_check.json path without mktemp and does not remove it. On shared multi-user systems, another local user can pre-create or tamper with that file to disrupt or spoof capacity checks.
Fix: Process ALL_REGIONS_JSON in memory by piping directly to jq (as done in workflow.md) or create a secure temporary file with mktemp and remove it on exit.
