How do you handle an unreliable or bad-actor agent today?

[ucsandman]

Hey, quick question, I'm not selling anything. I'm building a reputation oracle for AI agents (Sybil-resistant scoring, signed receipts, open source: https://github.com/ucsandman/Agent-Reputation-Oracle) and trying to figure out if it solves a real problem before I build more of it.

When an agent listed in the registry turns out to be unreliable or a bad actor, how do you handle that today? A blocklist, a manual review queue, something else?

Genuinely just trying to learn what already exists. Happy to be told this is a non-problem.

[Michael-WhiteCapData]

The official MCP Registry does not currently provide reputation scoring or a blocklist for agents. It is also important to distinguish that the Registry catalogs MCP servers, not AI agents.

For published servers, the current trust layer starts with namespace authentication, which verifies that a publisher controls the claimed GitHub account or domain. The Registry then applies a deliberately permissive moderation policy. Maintainers can manually mark malware, spam, illegal content, or completely broken servers as deleted, but they explicitly do not promise broad security review or removal of every buggy or vulnerable server.

Security scanning and reputation are intentionally left to package registries and downstream aggregators. The Registry documentation specifically describes aggregators adding ratings, security scans, and other custom metadata. That means a reputation oracle would not duplicate an existing Registry feature. It would fit more naturally as a downstream aggregator or subregistry that consumes Registry metadata and adds evidence based reputation signals.

If you take that route, I would key each result to the verified publisher namespace, server name, exact version, and package digest where available. That avoids carrying a good or bad score across unrelated releases and keeps signed receipts tied to the artifact that actually produced them.

The relevant documentation is in About the MCP Registry, the moderation policy, and the aggregator guide.