mcp-guard v0.4: Audit Any MCP Server in One Command, Detect Rug-Pulls and Test Prompt Injection
ErgoniaWorks · reddit · 2026-10-13
mcp-guard (MIT, Python) v0.4 shows what an MCP server actually does before you add it:
- Live audit, no config: one command starts the server, does the handshake, pulls the real tools/list and rates every tool. On the official memory server: the launch line rates LOW, but 6 of 9 tools can create or delete data — expected, yet a config line never tells you.
- Rug-pull detection: lock fingerprints every tool (description, schema, annotations) into mcp-guard.lock; commit it and inspect --lock flags tool-changed/added/removed when a server quietly mutates after approval. Secrets never enter the lock, so key rotations don't fire alerts.
- Automated prompt-injection runs: test-injection plays a hostile MCP server against your agent with honeypot tools and hidden payloads, failing on canary leaks (even base64/reversed), unneeded tool calls, or data exfiltration; includes a real-Claude adapter — run multiple times since results aren't deterministic.
Honest limits: pattern-based, reads declarations not runtime behavior — a tripwire, not a sandbox. v0.4 also tightens false-positive rules ("Execute a SQL query" no longer flagged as shell exec).
More from coding & agent
- Cloud vs On-Prem Is the Wrong Question for AI Agents: Split Workloads by Component — DavidLinthicum · 2026-10-13
- Keeping AI Agents Current as Business Rules Change: Versioned Rules and Event-Based Invalidation — tomibrumen · 2026-10-13
- Open-Source Python Coding Agent 'Agent Code' Ships with GUI and BYOK Support — Possible_Dinner_323 · 2026-10-13
- Using Claude computer use to drive ChatGPT: two AI agents reviewing code together — airkatakana · 2026-10-13
- RAG Support Bot Shines in Demos, Falls Apart on Real User Questions — Sad_Lavishness_53 · 2026-10-13
- Boris Cherny had 12 Claude Code agents turn 64 research papers into a browser lab you can rerun — ID_R_McGregor · 2026-10-13