Report privately to probes@agentcount.ai. Please do not open a public issue for a vulnerability.

Include what you can: what you did, what happened, and — if it is safe to — how to reproduce it. A rough report beats no report; you do not need a proof of concept.

Expect an acknowledgement within a few days. This is a small project with one maintainer, so please read that as a good-faith intention rather than a guaranteed SLA.

There are none, and that was verified rather than assumed. Before this repository was made public, every commit in its history was audited for committed .env files, RPC provider keys, database URLs and JWT-shaped strings. Nothing was found: the only connection strings that have ever appeared are localhost and an obvious user:pass@host/db placeholder in documentation.

If you find something that audit missed, that is exactly the kind of report this document is for.

The things most worth attacking here, roughly in order:

• The SSRF guard (crates/probe/src/netguard.rs). Every agent-supplied URI is hostile input by construction: anyone can register an agent and point its tokenURI anywhere. Requests are restricted to public IP addresses, and every redirect hop is re-validated, not just the first request. A way to reach a private, loopback, link-local or cloud-metadata address — including via DNS rebinding, a redirect chain, or an IPFS gateway rewrite — is the highest-severity bug in this codebase.
• The frontend's URL handling. Agent-supplied strings are rendered as links. data:, javascript:, vbscript:, file: and blob: are refused by a scheme allowlist, and third-party links carry rel="noopener noreferrer nofollow ugc". A bypass is in scope.
• SQL injection or unbounded resource use in the API (crates/api).
• Anything that lets a third party alter a stored run, which would break the immutability the whole method rests on.

• A check you disagree with. That is CONTRIBUTING.md's process, not a security report.
• A published number you believe is wrong. Genuinely welcome — open a public issue with the run id and agent id. This project has retracted its own findings before publication more than once.
• Volume or timing of our probe traffic. Email the same address and we will suppress a host or agent from future sweeps; see METHODOLOGY.md §6.

Coordinated. Tell us, give us a reasonable window to fix it, then publish whatever you like — including that we were slow, if we were. Fixes to check semantics that result from a report still go through the public process in CONTRIBUTING.md, because a check that changed for unexplained reasons is its own kind of problem.