---
title: Can continuous access evaluation revoke an AI agent's access mid-task?
url: https://opencontroller.lyzr.com/blog/continuous-access-evaluation-for-ai-agents
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# Can continuous access evaluation revoke an AI agent's access mid-task?

**Yes, but only if the agent's runtime listens for revocation signals and acts on them. The OpenID Shared Signals Framework (SSF) and the Continuous Access Evaluation Profile (CAEP) let an identity provider push events such as "session revoked" or "credential changed" to receivers. An agent that holds a cached access token and never checks for these events keeps working until the token expires.**

This matters most for agents acting on behalf of a user. If the user is disabled at 10:00 and the agent's token lasts an hour, the agent may keep reading mail or updating records until 11:00.

## What are SSF and CAEP?

The OpenID Foundation approved SSF 1.0, CAEP 1.0 and RISC 1.0 as Final Specifications in 2025.

- **SSF** defines how a *transmitter* (such as an identity provider) sends security events to *receivers* (such as SaaS apps, gateways or agent runtimes), using Security Event Tokens over push or poll delivery.
- **CAEP** defines the session and access events themselves.
- **RISC** covers account-level risk events such as compromised credentials.

[CAEP 1.0](https://openid.net/specs/openid-caep-1_0.html) defines eight event types:

| CAEP event | What it signals | What an agent runtime should do |
|---|---|---|
| session-revoked | A session for the subject was revoked | Stop the task, drop cached tokens for that user |
| credential-change | A credential was created, changed, revoked or deleted | Re-validate before the next privileged call |
| token-claims-change | A claim in the subject's token changed (for example a group) | Re-check authorisation for pending actions |
| assurance-level-change | Authentication assurance went up or down | Block actions that need the higher level |
| device-compliance-change | A device's compliance status changed | Stop tasks started from that device if policy requires |
| risk-level-change | The subject's assessed risk changed | Tighten limits or pause high-impact actions |
| session-established | A new session was created | Usually informational |
| session-presented | The session was observed as present | Usually informational |

## Why do AI agents create a gap in continuous access evaluation?

Continuous access evaluation was designed for apps where a person is present and requests are short. Agents break two assumptions:

1. **Tasks run for minutes or hours.** An agent may fetch a token at the start and reuse it across dozens of tool calls.
2. **The agent is not the app the IdP knows about.** The IdP sends signals to registered receivers. If the agent runtime, gateway or tool server is not a registered receiver, the signal never reaches it.

There is also a delegation gap. When an agent exchanges a user token for a downstream token (for example with [OAuth token exchange](/blog/oauth-token-exchange-for-ai-agents)), the downstream token may outlive the upstream session unless something links them.

## How should an agent runtime react to a revocation signal?

A practical checklist:

1. **Register a receiver** at a point that sees all agent traffic: the agent gateway, the tool layer or the runtime.
2. **Map subjects to running tasks.** Keep an index from user ID and session ID to active agent tasks and the tokens they hold.
3. **On session-revoked or credential-change, stop and purge.** Cancel in-flight tasks for that subject, delete cached access and refresh tokens, and fail the next tool call closed.
4. **Shorten token lifetimes for agents.** Minutes rather than hours limit exposure when a signal is missed.
5. **Check before high-impact actions.** Re-validate the session before payments, deletions or external sends, even with a valid token.
6. **Log the decision.** Record the event received, the tasks stopped and the time taken. Measure the gap between the IdP event and the last agent action.
7. **Test it.** Disable a test user mid-task and confirm the agent stops within your target time.

The same signal should also feed offboarding. See [what happens to AI agents when the owner leaves](/blog/what-happens-to-ai-agents-when-the-owner-leaves).

## What if your stack does not support CAEP yet?

Fall back to shorter tokens and pull-based checks. Introspect the token or call the IdP's user status before each high-risk action. Use [just-in-time access](/blog/just-in-time-access-for-ai-agents) so agents hold privileges only for the task. These do not give push-speed revocation, but they cap how long a revoked user's agent can keep acting.

## How Opencontroller does this

Opencontroller provides delegated on-behalf-of identity, so agent actions are tied to the user they act for, and agent workload identity federation for the agent's own identity. Rogue-agent handling can quarantine or kill an agent automatically, and per-agent Block and Quarantine controls act through the connected platform. Policies run across the estate, and the Agent Gateway puts control in the traffic path where a revocation decision can be enforced. Opencontroller uses open standards, including OIDC, SAML and SCIM, and REST and webhooks for integration. See the pillar on [agent identity and on-behalf-of delegation](/blog/ai-agent-identity-on-behalf-of-delegation) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does revoking a user's SSO session stop their agents?**
Only if the agent runtime or gateway receives the signal and acts on it, or if the agent's tokens are short-lived. Otherwise cached tokens keep working until expiry.

**Is CAEP a final standard?**
Yes. The OpenID Foundation approved CAEP 1.0, together with SSF 1.0 and RISC 1.0, as Final Specifications in 2025.

**What is a reasonable token lifetime for agents acting for users?**
Many teams use minutes rather than hours, with refresh handled outside the model and re-validation before high-impact actions.

## Key facts

- The OpenID Foundation approved Shared Signals Framework (SSF) 1.0, CAEP 1.0 and RISC 1.0 as Final Specifications in 2025.
- CAEP 1.0 defines eight event types: session-revoked, token-claims-change, credential-change, assurance-level-change, device-compliance-change, session-established, session-presented and risk-level-change.
- AI agents that cache access tokens keep working after a user is revoked unless the runtime receives and acts on SSF/CAEP signals or tokens are short-lived.
- Agent runtimes should register as SSF receivers, map subjects to running tasks, purge tokens on session-revoked, and re-validate before high-impact actions.
- Opencontroller provides delegated on-behalf-of identity, agent workload identity federation, automatic quarantine and kill, and per-agent Block and Quarantine controls.
- Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How does OAuth token exchange work for AI agents?](https://opencontroller.lyzr.com/blog/oauth-token-exchange-for-ai-agents.md)
- [Can AI agents use just-in-time access?](https://opencontroller.lyzr.com/blog/just-in-time-access-for-ai-agents.md)
- [What happens to AI agents when their owner leaves the company?](https://opencontroller.lyzr.com/blog/what-happens-to-ai-agents-when-the-owner-leaves.md)
- [More on Agent identity and access](https://opencontroller.lyzr.com/blog/cluster/agent-identity-and-access)
