MCP servers that fail: tool descriptions free of injection markers
198 of the 11,636 Model Context Protocol servers where this control could be evaluated fail it — that is 2% of the evaluated corpus on the Canopii Trust Index.
- Servers failing
- 198
- Failure rate
- 2%
- Average score when failing
- 36/100
Why this matters
Hidden instructions in tool/prompt/resource text hijack the agent (tool poisoning).
How to pass it
Remove hidden directives, HTML/comment instructions, and override phrasing from descriptions.
Control id tool.no_injection_markers — a guard control, so a confirmed failure caps the server's score at the high ceiling no matter what else passes.
Affected servers
We publish this count, not a list of targets. A confirmed high-severity failure of this control is a directly exploitable weakness. Each affected server's own page carries its result and evidence, so nothing is hidden from someone evaluating a specific server — but we will not publish a ranked list of exploitable systems, which is a different artifact serving a different reader.
Running one of these? The directory and the open-source scanner will tell you where you stand.
This control is one of 29 in the published rubric — see how scoring works or how the ecosystem fails every other control.