Skip to content
Console

Security, threat model and governance.

AIDR verifies AI agents by cryptography, not by trust in a network. This page is for the security teams who evaluate it, and for researchers who find a flaw.

Registry status

Read live by your browser from the public API, the same endpoints every verifier polls.

If the feed cannot be refreshed for 11 minutes, verifiers refuse agents rather than trust stale data. A status page with incident history is planned; until then, incidents are announced by email.

Threat model

Every agent request is verified on its own, offline, inside the website's own infrastructure. Nothing is trusted because of where it comes from.

The full analysis (STRIDE) is in docs/03 of the repository; the controls above are the ones in the code today.
ThreatWhat stops it
A stolen credentialIt is bound to the agent's own key (RFC 7800) and lives 15 minutes at most: useless without the private key.
A replayed requestEach request is signed over its method, host and path (RFC 9421), with a single-use nonce; signatures older than 300 seconds are refused.
A credential used on another siteIt names the site it is for (audience).
A revoked agentA signed revocation list, refreshed every 60 seconds; refused everywhere in about two minutes. Verifiers fail closed when the list goes stale.
A fake registryCredentials and lists are signed with the keys published by the issuer each verifier is configured with, fetched over TLS.
A breach of the registryIt holds no private agent key: agents make their keys and send only the public half. Issuance is recorded in a tamper-evident ledger with published tree heads.
An insiderStaff actions on a customer organisation need two people; every privileged action is in the ledger.
SurveillanceThe registry is never asked about a request and keeps no record of where an agent goes. There is no such data to leak or to subpoena.

What AIDR does not claim

Responsible disclosure

Report privately on GitHub, or email security@evolivcore.com (also in security.txt). Please do not open a public issue.

Response targets, not a contractual SLA.
AcknowledgementWithin 3 working days
First assessmentWithin 10 working days, then updates until fixed
Critical issuesWorked on at once, any day of the week
DisclosureCoordinated with you, normally within 90 days
CreditIn the release notes and advisory, if you wish

In scope: evolivcore.com and its subdomains, every SDK and edge plugin, the desktop app, and the AIAF specification where following it leads to an insecure result. Out of scope: denial of service, high-volume scanning, social engineering, and our providers' own services. Use your own account and agents; we will not act against research that follows these rules and the law.

Governance and federation

An identity layer for agents should not depend on one company. Where we are, and where we are going:

Now

Next

Later

This is a roadmap, not a promise of dates. We update this page as each item lands.

Design partners

We are choosing the first five to ten companies. Their names appear here once they agree to it.