MCP-Guard 0.14.0 boosts tool-poisoning detection from 10/15 to 15/15, blind spots remain
ErgoniaWorks · reddit · 2026-11-14
The author released MCP-Guard 0.14.0, focusing on making MCP tool-poisoning detection more accurate and interpretable.
- On a historical synthetic benchmark, detection improved from 10/15 to 15/15 attacks caught with zero false positives across 34 benign examples; on a separate challenge set the detector still missed 4 of 6 attacks and produced one false positive.
- New confidence levels for capability detection and an experimental --combined analysis flag which risk capabilities are confirmed, missing, or uncertain.
- Static scan reads local MCP configs without starting servers: uvx mcp-guard-audit scan; live inspection is also available.
MIT-licensed, on PyPI; the author is soliciting feedback on false positives and missed detections in real setups.
More from coding & agent
- Hermes AI Agent Night at AWS Singapore on Dec 16 to preview Hermes mobile app — Teknium · 2026-11-14
- Running AI Agents in Production: Reddit Compares 6 Platforms' Pricing Models — AcanthaceaeLatter684 · 2026-11-14
- Agentic formal verification tool vstack finds 17 bugs in vLLM, fixes underway — AccBalanced · 2026-11-14
- Openship: self-hosted deployment platform with built-in CI/CD hits 15.3k GitHub stars — jackedAJ · 2026-11-14
- Money-saving trick: let a cheap model do the work, then ask it to summon a pricier model to review — vista8 · 2026-11-14
- Open-source personal agent The Machine drafts everything but never sends without your OK — markloperman · 2026-11-14