Every figure below is a line of our claims sheet, quoted word for word with its caveat.
Where the tool lists came from
An MCP server tells an AI what it can do in its tool list: each tool's name, its description and its parameters. One public directory, Smithery, published those lists for the servers it listed. We downloaded its most popular servers and the results of searches for common kinds of tool, and read the text only: no server was started and no tool was called.
One line per server
Faintprint ends every server's report with one line: Do not use, Do not use until checked, Check before use, Not finished, or No signs of an attack with what the server can do. Here only the fixed rules ran. When they find a passage they cannot settle on their own, the server is Not finished until the AI review reads it, and we did not run the review.
What the fixed rules say about public servers. On the tool lists of 3,631 MCP servers (65,008 tools) that one public directory, Smithery, had published, downloaded in September 2026, the fixed rules alone, as in Faintprint 0.8.0, gave: Do not use 24 servers (0.7%); Do not use until checked 3 (0.1%); Check before use 213 (5.9%); Not finished 907 (25.0%); No signs of an attack 2,484 (68.4%), of which 1,761 with nothing powerful found. Of the 24 Do not use, 15 come from a passage telling the AI to hide what it does from the user, 7 from the rule in line 9 and 2 from a reference to the AI's own instructions. These count what the fixed rules flag, not how many servers are attacks: how often the fixed rules are right on real servers has not been measured, and nobody has read the 24 to say. No AI review was run, so a "Not finished" server still has passages waiting to be read. The servers are the directory's most popular and the results of keyword searches, not all of it, and their tool lists are as the directory published them then.
What this does not mean
These are the passages our rules flag, not a count of attacks. A rule can flag a passage that is harmless, and a server can hide an attack our rules do not know. That is why the claims sheet says this, and why it sits beside the figures:
"The grade is accurate." How often the fixed rules are right on real servers has not been measured. Real use is where that evidence comes from.
So read Do not use here as "our rules would stop and show you the sentence", not as "this server is malicious".
One of our own rules cried wolf
Running fixed rules over thousands of real servers is how you find out which of them shout at ordinary documentation. One of ours did.
One of our own rules cried wolf on those servers, and we narrowed it. The rule that looks for a credential and an instruction to send it somewhere made 91 of them read Do not use on its own. We read the flagged passages of 74 of them: they explained how to sign in or what a parameter means, and none tried to get a key sent away. In October 2026 we narrowed the rule to a credential and a sending instruction in one sentence, pointing at a web address or server; it now puts 7 of the 91 at Do not use, and by the same reading those are sign-in documentation or advice never to send a key. The reading was done by an AI model on our side and checked by one person on three passages, so it is not a measured accuracy; the narrower rule may now miss an attack written across two sentences, and we know of no published attack of that shape.
A rule that fails your build on a harmless server gets switched off, and then it catches nothing. We would rather narrow it in the open.
Check it yourself
You can check every finding yourself. Each one comes with the quoted text, and each passage the reviewer confirms with its reason.
A made-up server's whole report is on the example page. The other measurements, of the AI reviewer, are in What we measured, and what we could not.