Fixed-rule scan of 3,631 public MCP servers flags 24 as do-not-use
Minimum-Gift-5217 · reddit · 2026-11-20
The author of Faintprint ran its pure fixed-rule engine (no AI, no API key) over the tool lists of 3,631 public MCP servers (65,008 tools) published by the Smithery directory. Results: 24 servers (0.7%) rated do-not-use — 15 for passages telling the AI to hide its behavior from the user — plus 213 check-before-use and 907 not-finished; 2,484 showed no signs of attack.
They also disclose their own false-positive lesson: a rule matching credential-plus-send-instruction alone flagged 91 servers, but reading 74 of them showed all were sign-in docs; narrowing the rule to one sentence pointing at a web address cut it to 7. The authors stress these are rule flags, not measured attack counts, and no servers are named.
More from coding & agent
- Karpathy's 70-minute breakdown of how the top 1% actually work with LLMs — blaizedsouza · 2026-11-20
- Karpathy: prompting is going away — graphs are the only layer that survives — blaizedsouza · 2026-11-20
- Obsidian Starter Kit v6 bundles your vault into one file for Claude Code with token limits — dSebastien · 2026-11-20
- Browser MCP benchmark: 176k-565k tokens per bug fix across 5 tools — VisionInBinary · 2026-11-20
- Developer uses Enhanced Radar MCP to monitor his own flight's ATC comms live — massimosgrelli · 2026-11-20
- tinyeval: a Bash-only minimal eval framework with a Claude Code skill — criticasterdotcom · 2026-11-20