Notes
Notes
What we measured, and how to check the MCP servers your AI tools use.
- What tool poisoning looks likeOne poisoned tool, from a server we made up: what it asks the AI to do, the three things that give it away, and what Faintprint says. A guide, November 2026.
- What our fixed rules flag on public MCP serversOur fixed rules over the tool lists a public directory had published: what the one line looks like across real servers, what it does not mean, and one of our rules caught crying wolf. November 2026.
- What we measured, and what we could notHow often our AI reviewer is wrong, on harmless servers, on published attacks and on attacks aimed at the reviewer itself, each figure with its caveat. November 2026.