How to revoke an AI agent's credential and stop its access
Revoking a credential and stopping an agent are different controls. The procedure, caches, permission withdrawal and tests with measured times.
Revoking an AI agent's credential communicates that the identity should no longer be accepted. Stopping its activity also requires withdrawing permissions, cancelling jobs and checking that it cannot carry on acting. They are related operations, with different effects and different timings.
The URL keeps its historical wording, "in seconds". We do not publish a guaranteed cut-off latency: it has to be measured in the real environment, including caches, sessions and connectors.
Revoke, expire and stop
| Action | What changes | What it does not prove |
|---|---|---|
| Expire | The declaration passes its validity date | That existing sessions are cancelled |
| Revoke | The public status says it should not be accepted | That every connector has withdrawn access |
| Withdraw permissions | Effective access is reduced or removed | That the public credential has been updated |
| Cancel execution | A job or session is interrupted | That no reusable tasks or tokens remain |
| Reissue | A new approved declaration is signed | That the previous version should still be accepted |
CORTEX uses a signed status list. The lifecycle documentation explains the mechanism. The status each consumer sees depends on its queries and on caches.
A withdrawal procedure you can check
- Identify the agent, the credential, the jobs in progress and the connectors affected.
- Contain the activity: withdraw the appropriate access and cancel jobs, sessions and queues.
- Publish the revocation of the credential, recording the reason, the responsible person and the time.
- Query the verifier from outside and record when the status changes.
- Attempt a harmless test action in each connector and confirm it is rejected.
- Check that there is no alternative route using other credentials.
- Keep the evidence and decide whether a new issuance is appropriate.
A visual seal can still show up in a screenshot, an email or a cache. The image is not a switch and it is not a reliable source of current status.
How to measure it without putting operations at risk
The test should use an isolated test agent and fictitious data. Record one permitted action, one outside scope and a third after withdrawing permissions and revoking. Keep the authorisation decision and the connector's outcome, along with the timings.
At minimum, measure: time of decision, publication of the status, first query that detects the revocation, effective withdrawal of access, and the last action attempt that was accepted. Reporting a single figure without saying which interval it measures can mislead the buyer.
We have not revoked Argos to illustrate this article. Its public verification lets you consult the credential; it is not proof that all of its integrations have been stopped.
What happens when the configuration changes
The hash compares the registered manifest with the signed reference. It does not automatically observe a prompt loaded into another service, a change of model, or permissions granted directly in a tool. The procedure has to detect and approve those changes, check the effective configuration and reissue where appropriate.
The scope of the issuance and the signed prohibitions are reviewed separately from the manifest. Human approval does not mean permission to ignore a prohibition.
Frequently asked questions
How long does a revocation take?
It depends on the service publishing the status, on caches and on each consumer. The operational cut-off adds its own timings. It has to be tested and measured.
Can a revoked credential still be used?
An integration that does not check the status, or that uses an old response, could accept it. That is why the freshness policy and the behaviour on errors matter.
Does the seal prevent an attack?
No. It makes a declaration easy to check; prevention and containment require other controls.
Review the digital ID and its limits or ask us to review the control of your agents.