Responsible disclosure
MCP Hub audits the ecosystem to protect AI-agent users. Our goal is to build trust, not to name and shame. Here is how we publish our signals.
Our approach
- We raise a security signal on a server and present it with our analysis (what we observed, which category, why).
- The server author has a right of reply: they can annotate the signal ("legitimate", "fixed", "disputed") and add a note.
- We do not erase the signal: it stays displayed, next to the author's response. The dialogue is public.
- When a server passes the behavioral pentest, that verdict is what counts in the directory — not the mere description analysis.
Our commitments
- A server's code never runs outside our isolated sandbox (micro-VM, controlled network).
- Our pages publish names and counters — never the raw content returned by a tool.
- A description analysis is not a verdict: it is indicative; only observed behavior decides.
- We favor measurement and proof over accusation: a signal is an invitation to verify and respond.
Are you the author of a flagged server?
You can exercise your right of reply and, if needed, request a re-run after fixing. The goal is for your server to show the most accurate, up-to-date verdict possible.