Why doesn’t an AI kill switch stop an agent’s access everywhere at once?

Why doesn’t an AI kill switch stop an agent’s access everywhere at once?

An AI kill switch usually stops the agent itself, its process, session, or runtime. The access the agent collected along the way lives elsewhere, in tokens and grants held by every app it connected to. Pulling the switch ends the agent’s next decision, not the credentials it already holds.

What an AI Kill Switch Stops and What Keeps Running

Most kill switches act at runtime. They terminate the agent process, end its platform session, or block its calls to the model. That stops new reasoning, which matters, but an agent’s reach rarely ends at runtime.

Several things persist after the switch. Access tokens issued to the agent continue working until they expire, and refresh tokens stored in a connector can mint new ones. Jobs the agent started in other systems, webhooks it registered, and keys it created continue operating according to their own schedules. A sync job the agent set up on Monday will still run on Tuesday, kill switch or not. Pluto’s guidance on enterprise agent governance describes containment as three actions together: “Revoking agent credentials, terminating active sessions, blocking tool access.” A kill switch that only ends the session leaves the other two undone.

Why Issued Tokens Stay Valid After the Kill Switch Fires

Most OAuth access tokens are bearer tokens, so anyone who holds one can use it. Many are self-contained, and the receiving API validates them locally by checking the signature and expiry. RFC 7009, the OAuth standard for token revocation, notes that with such tokens “a resource server needs no further interaction with an authorization server.” If nothing tells that API that a token was revoked, it keeps accepting it.

The same RFC requires authorization servers to support revoking refresh tokens but only recommends supporting revocation of access tokens. When that support is missing, it states that “access tokens will not be immediately invalidated when the corresponding refresh token is revoked.” One major identity platform assigns access tokens a default lifetime of 60 to 90 minutes. That window applies to every token the agent already holds and to every API that accepts them.

How Delegated OAuth Access Keeps Agents Connected to Other Apps

Agents rarely run on a single credential. When a user connects an agent to email, a code host, or a CRM, each app’s authorization server records a consent grant with its own OAuth scopes and issues its own tokens. That delegated access belongs to the grant, not the agent’s runtime, so disabling the agent on one platform doesn’t affect a grant recorded elsewhere.

Refresh tokens make those grants durable. One major provider documents that its refresh tokens remain valid until the user revokes the app’s access or the token goes unused for six months. Broad scopes widen the damage, because a grant for full mailbox or repository access outlives the task it was approved for.

Where credentials are stored matters too. Pluto’s review of Claude Managed Agents found that vault credentials never enter the agent’s sandbox, which is a strong default against theft. Stopping a sandbox, though, is not the same as revoking what the vault holds.

Revoking an Agent’s Access Everywhere During an Incident

Effective AI incident response works from the agent’s grants outward rather than from its process inward.

  • Inventory: List every app, scope, and credential the agent holds, so revocation has a target list.
  • Revoke at the source: Remove refresh tokens and consent grants at each authorization server, then rotate any static keys.
  • Close the window: Keep access token lifetimes short, or have APIs check tokens through introspection so revocation lands immediately.
  • Push the signal: Use the OpenID Foundation’s Shared Signals Framework and CAEP, both final since September 2025, to tell connected apps a session has ended.

Keep the evidence. Pluto’s review notes that deleting a Claude Managed Agents session is “a hard delete with no audit trail,” while archiving keeps the event log.

Final Thoughts

An AI kill switch is still worth having. It stops the agent from making its next bad decision and buys time. It just doesn’t reach the access the agent has already spread across other systems. Treat the switch as the first step of incident response, not the whole process, and design delegated access so revoking it everywhere is a list you can work through in minutes.