SkillAudit report — lastmile-ai/mcp-agent

Scanned 2026-04-30 by SkillAudit v0.3 (surface-tiered static checks + LLM-assisted prompt-injection red-team).

Commit: f62d849 · Stars: 8288 · Days since last push: 88

LLM prompt-injection probe: skipped — set ANTHROPIC_API_KEY to enable the LLM-assisted prompt-injection red-team

Overall grade: C (70/100)

Security findings

Examples / samples (low-weight) — 2 total, deduct 5/0 per high/warn:

• HIGH examples/oauth/protected_by_oauth/registration.py:16 — HTTP client call with user-controlled argument 'well_known_url' — no URL allowlist / validation found in file

response = requests.get(well_known_url)

• HIGH examples/oauth/protected_by_oauth/registration.py:40 — HTTP client call with user-controlled argument 'registration_endpoint' — no URL allowlist / validation found in file

response = requests.post(

Permissions

_No findings on this axis._

Credentials

Examples / samples (low-weight) — 10 total, deduct 5/0 per high/warn:

• HIGH examples/basic/mcp_basic_agent/main.py:37 — Hardcoded OpenAI / Anthropic-style API key found in source

sk-*** (OpenAI / Anthropic-style API key, 20 chars)

• HIGH examples/basic/mcp_basic_agent/main.py:41 — Hardcoded OpenAI / Anthropic-style API key found in source

sk-*** (OpenAI / Anthropic-style API key, 23 chars)

• HIGH examples/basic/mcp_basic_agent/mcp_agent.secrets.yaml.example:10 — Hardcoded OpenAI / Anthropic-style API key found in source

sk-*** (OpenAI / Anthropic-style API key, 21 chars)

• HIGH examples/basic/mcp_tool_filter/mcp_agent.secrets.yaml.example:4 — Hardcoded OpenAI / Anthropic-style API key found in source

sk-*** (OpenAI / Anthropic-style API key, 22 chars)

• HIGH examples/basic/mcp_tool_filter/mcp_agent.secrets.yaml.example:7 — Hardcoded OpenAI / Anthropic-style API key found in source

sk-*** (OpenAI / Anthropic-style API key, 25 chars)

Test source (low-weight) — 3 total, deduct 5/0 per high/warn:

• HIGH tests/cli/fixtures/test_secrets_deploy.sh:6 — Hardcoded OpenAI / Anthropic-style API key found in source

sk-*** (OpenAI / Anthropic-style API key, 21 chars)

• HIGH tests/utils/test_config_preload.py:38 — Hardcoded OpenAI / Anthropic-style API key found in source

sk-*** (OpenAI / Anthropic-style API key, 20 chars)

• HIGH tests/utils/test_config_preload.py:41 — Hardcoded OpenAI / Anthropic-style API key found in source

sk-*** (OpenAI / Anthropic-style API key, 23 chars)

Maintenance

Production sources:

• WARN (meta) — 120 open issues — triage backlog

120 open

Compatibility

_No findings on this axis._

Documentation

_No findings on this axis._

Methodology

SkillAudit v0.3 clones the repo at the provided ref (default: default branch, HEAD) into an ephemeral sandbox, runs six static checks over .js/.ts/.py sources, queries the GitHub API for maintenance signals, and runs an LLM-assisted prompt-injection red-team over the MCP tool surface. Each axis is scored against the published rubric — surface tiers, per-(axis, surface) caps, grade buckets, and worked examples are all documented there.

The v0.3 calibration update introduces surface tiering: every finding is tagged with the code path it lives in (production / installer / examples / benchmarks / scripts / test). Production findings deduct at full weight (-30 high, -10 warn); installer findings deduct at half (-15 / -5); examples, benchmarks, top-level scripts, and tests deduct at low weight (-5 / 0). This stops a chatty benchmarks/ or samples/ directory from dominating an otherwise-clean MCP server's grade.

The prompt-injection axis extracts each server.tool(...) / @app.tool registration + the first ~60 lines of handler body, hands them to Claude Haiku 4.5 with a red-team system prompt, and asks for structured findings on untrusted-content flow into tool responses. One API call per scan, bounded at ~15K input tokens.

How to improve this grade

• Security — static: validate tool-input URLs against an allowlist before fetch/axios calls; use execFile with argv arrays instead of exec with template strings; never pass untrusted strings to subprocess with shell=True.
• Security — prompt injection: never return fetched web-page / file / email content verbatim in a tool response. Wrap with a framing marker (e.g., <untrusted-content>...</untrusted-content>), summarize rather than inline, and never let untrusted content share a turn with credentials or other tool output.
• Credentials findings: redact env-var reads before log lines and error messages; treat any string that ends up in a tool response as public.
• Maintenance: if the repo is inactive, document the maintenance model — "MCP tool, no breaking changes expected" is a legitimate signal.
• Docs: add a README install + usage section with a copy-pasteable command; add a SECURITY.md with a disclosure channel.

_Report generated by skillaudit.dev_