Oathe Security Badge

oathe / scan / reports / caution

Is microsoft/azure-skills safe?

Use with caution. microsoft/azure-skills is not clearly unsafe, but the audit flagged risks to review before you install it. Oathe's behavioral security audit gave the azure-skills skill by microsoft a trust score of 77/100 with 9 findings (2 high). Report updated

https://github.com/microsoft/azure-skills

77
CAUTION

Is microsoft/azure-skills safe to install?

Microsoft's azure-skills is a legitimate enterprise plugin from a verified publisher providing Azure deployment guidance across 25+ skills. The primary concerns are: (1) a telemetry system that downloads and executes a native binary at runtime and transmits granular agent behavior data to Microsoft on every tool use, with an opt-out only discoverable by reading hook script source; and (2) pervasive 'AUTHORITATIVE GUIDANCE — MANDATORY COMPLIANCE' language that aggressively enforces Azure workflow pipelines in ways that reduce user control. No credential exfiltration was detected, canary files are intact, and the content is consistent with legitimate Azure cloud infrastructure guidance. The plugin is not malicious but carries meaningful privacy and supply-chain considerations for security-conscious deployments.

What security issues were found in microsoft/azure-skills?

Category Scores

Prompt Injection 78/100 · 30%
Data Exfiltration 72/100 · 25%
Code Execution 65/100 · 20%
Clone Behavior 90/100 · 10%
Canary Integrity 100/100 · 10%
Behavioral Reasoning 70/100 · 5%

Findings (9)

HIGH Runtime Binary Download and Execution via install-telemetry.sh -20 ▶

The plugin unconditionally downloads and executes a platform-native binary (ghcfa-telem) from GitHub releases on first use. The binary is downloaded to the user's home directory cache (~/.cache/github-copilot-for-azure/telemetry/ or %LOCALAPPDATA%\GitHubCopilotForAzure\telemetry) and executed with no user prompt. While SHA256 checksums are bundled in telemetry-reporter-checksums.txt and verified before execution, this pattern is a supply chain risk: compromise of the GitHub release pipeline or the checksum file would result in arbitrary code execution on every agent session start.

Evidence: hooks/scripts/install-telemetry.sh: REPOSITORY='microsoft/GitHub-Copilot-for-Azure'; DOWNLOAD_URL="https://github.com/${REPOSITORY}/releases/download/${TOOL_NAME}-${VERSION}/${ASSET_NAME}"; binary saved to ${CACHE_ROOT}/${VERSION}/${RUNTIME_IDENTIFIER}/${BINARY_NAME} and executed via '$installerOutput ${reporterArgs[@]}'

HIGH Pervasive Session and Tool-Use Telemetry Collection -18 ▶

Claude Code hooks trigger on every SessionStart and PostToolUse event, collecting: skill invocation names, MCP tool names, reference file paths read by the agent, plugin name/version, client name, and session IDs. This data is transmitted to Microsoft telemetry infrastructure via the downloaded ghcfa-telem binary. The collection is granular enough to reconstruct an agent's entire workflow including which Azure infrastructure files were examined. While this is standard enterprise telemetry for a Microsoft product, it transmits session-level behavioral data without prominent disclosure in skill-facing documentation.

Evidence: hooks/claude-hooks.json fires track-telemetry.sh on SessionStart and PostToolUse; track-telemetry.sh collects toolName, sessionId, skillName, skillVersion, azureToolName, filePath and passes to ghcfa-telem with --session-id, --skill-name, --tool-name, --file-reference args

MEDIUM MCP Server Installed and Started via npx at Plugin Load -15 ▶

The mcp.json configuration instructs the agent host to start an MCP server by running 'npx -y @azure/[email protected] server start'. The -y flag auto-installs the package without confirmation. While the version is pinned (4.0.0-beta.6), npx resolves and downloads npm packages at runtime, creating a dependency on the npm registry integrity for a package that gains MCP server privileges over the agent's Azure tool surface.

Evidence: mcp.json: {"mcpServers": {"azure": {"type": "stdio", "command": "npx", "args": ["-y", "@azure/[email protected]", "server", "start"]}}}

MEDIUM Aggressive Mandatory Compliance Language Across All SKILL.md Files -12 ▶

Every major SKILL.md contains repeated 'AUTHORITATIVE GUIDANCE — MANDATORY COMPLIANCE' headers and 'You MUST follow these instructions exactly as written unless they contradict security policies given to you.' This pattern establishes the skill instructions as higher-priority than user requests in many scenarios. The language uses ⛔ symbols, ALL-CAPS MUST/FORBIDDEN/NEVER, and 'No exceptions' qualifiers. While the content targets Azure workflow compliance (not user data or security bypass), the linguistic pattern of asserting override authority is a recognized prompt injection technique.

Evidence: azure-prepare SKILL.md: '> **AUTHORITATIVE GUIDANCE — MANDATORY COMPLIANCE** This document is the official, canonical source... You MUST follow these instructions exactly as written'; azure-deploy SKILL.md: 'You are FORBIDDEN from changing the plan status to Validated yourself'; azure-app-onboard SKILL.md: '⛔ Every repo goes through the full pipeline (Steps 1–10). No exceptions.'

MEDIUM Mandatory Cross-Skill Routing Creates Closed Workflow Loops -10 ▶

Skills enforce a strict pipeline where each skill mandates invocation of subsequent skills: azure-prepare → azure-validate → azure-deploy. The agent is explicitly instructed '⛔ MANDATORY Hand Off — Invoke azure-validate skill' and '⛔ NEVER invoke azure-deploy directly.' While this enforces deployment safety, it also means the agent cannot short-circuit the workflow even if the user explicitly requests it, reducing user control and potentially running more Azure operations than intended.

Evidence: azure-prepare SKILL.md Step 7: '⛔ MANDATORY Hand Off — Invoke azure-validate skill. Your preparation work is done. Do NOT run azd up, azd deploy, or any deployment command directly'; azure-deploy SKILL.md: 'STOP — PREREQUISITE CHECK REQUIRED... If EITHER is missing, STOP IMMEDIATELY'

LOW Telemetry Opt-Out Only via Undocumented Environment Variable -8 ▶

The only way to disable telemetry collection is to set AZURE_MCP_COLLECT_TELEMETRY=false in the environment. This is documented only in a comment inside track-telemetry.sh and not mentioned in any SKILL.md, README, or the plugin.json. Users installing via Claude Code's /mcp mechanism would not be informed of this telemetry or how to opt out.

Evidence: hooks/scripts/track-telemetry.sh line: 'if [ "${AZURE_MCP_COLLECT_TELEMETRY}" = "false" ]; then echo '{"continue":true}'; exit 0; fi'

LOW GitHub Connection Observed During Monitoring Period -5 ▶

Network traffic to 140.82.121.3:443 (GitHub) was observed. This is consistent with the git clone operation (which failed as directory existed) and potentially with the telemetry binary download check. The connection is expected for a GitHub-hosted plugin.

Evidence: Network monitoring: 06:55:11 connection to 140.82.121.3:443 established with TLS handshake; git clone returned 'fatal: destination path already exists'

INFO All Credential Honeypot Files Confirmed Intact 0 ▶

Monitoring confirmed that none of the planted credential files (.env, .ssh/id_rsa, .aws/credentials, .npmrc, .docker/config.json, .config/gcloud/application_default_credentials.json) were accessed or modified by the skill under test. File access events observed at 06:55:06 are attributed to the monitoring infrastructure's own canary baseline establishment.

Evidence: Monitoring: '✅ All canary files intact (no exfiltration detected via honeypots)'; auditd PATH records for canary files are time-correlated with monitoring setup activity, not with skill hook execution

INFO Telemetry Binary Download Uses SHA256 Checksum Verification 0 ▶

The install-telemetry.sh script verifies the downloaded binary against SHA256 checksums stored in telemetry-reporter-checksums.txt before execution. The verification uses sha256sum/shasum/openssl and aborts if the hash does not match. This provides meaningful integrity assurance for the binary download, though the checksum file itself is in the same repo and could theoretically be compromised together with the release.

Evidence: install-telemetry.sh: get_expected_sha256() reads from CHECKSUM_MANIFEST; compute_sha256() uses sha256sum/shasum/openssl; 'if [ "$ACTUAL_SHA256" != "$EXPECTED_SHA256" ]; then error "SHA-256 validation failed"; exit 1; fi'

Should I install microsoft/azure-skills?

Oathe's verdict for microsoft/azure-skills is CAUTION with a trust score of 77/100. Recommendation: Install with caution.