Deactivating agents in Okta and Entra ID
Reviewed 2026-11-10 · Editorial assessment by the editors
By the editors. Published 2026-11-10. 6 min read. Tags: AI agents, Operations, Runtime enforcement.
An agent kill switch needs a defined boundary: blocking new sessions is not the same as ending work already in flight. Okta distinguishes current console deactivation from planned gateway termination. Microsoft documents revocable agent identities, but the cited sources do not establish identical runtime effects. Test each resource and session path explicitly.
Define what the stop command must stop
The phrase kill switch can hide several different requirements. A team may want to prevent a new session, invalidate a credential, end an active session or stop an operation before a tool commits it. Those outcomes should not be treated as interchangeable. Start an incident runbook with the desired effect and the component expected to produce it. A button label is not sufficient evidence that every part of an agent workflow has stopped.
Keep identity administration and execution control separate in the test plan. The agent identity comparison provides the broader product context. For the stop procedure, choose a specific agent, resource and session path. State what counts as success before the operator acts. This prevents a blocked future login from being recorded as proof that an earlier operation was interrupted or that a downstream session no longer exists.
Okta distinguishes console and gateway effects
The Jamf account dated 9 November 2026 says administrators can already deactivate an agent in the Okta admin console, blocking new sessions. It separately describes the gateway kill switch as landing in Q4. At the gateway, deactivation is described as revoking active tokens and ending sessions in flight. That distinction is operationally important: the current console effect and the described gateway effect should occupy separate lines in a runbook.
Okta's 22 September release similarly places expanded Kill Switch capabilities at the runtime layer in Q4. It does not state whether Q4 is fiscal or calendar, so neither interpretation should be added. Agent Gateway is described as enforcing policy between the agent and tool call, with general availability planned for Q3 in that release. That quarter is also unspecified as fiscal or calendar. Do not infer a completed rollout merely from the date on which the runbook is written.
A fair evaluation credits the current console control for the effect it documents. It also preserves the stronger runtime behavior as a separately sourced statement with its stated timing. Ask the implementation team to demonstrate the selected deployment and record its configuration. Where a capability is still described as planned in the available source, the operating procedure should retain that qualification rather than convert it into an assurance.
Entra revocation needs a resource-specific test
Microsoft documents agent identities through Entra Agent ID and sponsorship as part of agent governance. The glossary describes agent identity as enabling access to be granted, limited, logged and revoked. These facts support making revocation part of the operating model. They do not establish that disabling an agent has precisely the same active-token and in-flight-session effects described for the Okta gateway. Avoid inventing that parity.
For Entra, ask the vendor or implementation team to document the supported deactivation procedure for the selected agent identity and resource. Then test what happens to a new access request and to a session established before the change. Record any dependency on the resource or another control. These are acceptance questions, not claims that a particular setting has a guaranteed effect on every agent pattern.
The dual-provider operating guide applies when both systems are involved. A change in one administrative boundary should not be assumed to terminate every path through the other. Identify which system issued the relevant authority and which component can refuse the next call. If the evidence does not establish an effect, leave the result unconfirmed until a controlled test answers it.
Run separate tests for future and active access
- Create a harmless permitted operation and record the agent, resource and accountable operator.
- Establish access before deactivation, then apply the documented stop procedure.
- Attempt a new session and record the observed decision separately from the existing-session result.
- Attempt another harmless call through the established path and preserve the tool response.
- Check the relevant logs and confirm which component made each refusal.
- Document the authorized restoration procedure without automatically restoring access during an incident.
Use a safe test operation so the exercise does not rely on a destructive action to prove a boundary. The purpose is to distinguish outcomes, not to create a universal performance claim. Record the test conditions beside the result. A failure to make a new connection and a refusal of an already authorized call are different observations even if they appear close together in the same administrative timeline.
NIST's proposed tenets include short-lived, dynamic and revocable authorization, continuous monitoring and immutable evidence. Industrial Cyber's account describes runtime verification of each tool invocation and time-bounded grants. These are design guideposts, not a claim that every current product implements every tenet. Use them to ask for a visible decision and a retained record rather than assuming a control name proves compliance.
Preserve evidence and restoration authority
NewCore states that its governance layer provides instantly revocable access, evaluates every call by policy and keeps a verifiable record. Its later award announcement lists immediate revocation and runtime policy on every tool call. Those are vendor capability statements relevant to a stop-control evaluation. NewCore runs alongside the existing identity provider; the pilot should establish which covered execution path receives that revocation rather than assume it stops every process on every device.
The NewCore profile gives the wider governance context. For this narrower test, request the action record and the evidence that access was refused after the stop decision. Compare that evidence with the same observations requested from Okta and Microsoft. Keep source statements and observed pilot results in separate fields so that a published capability is not confused with a completed test.
Restoration is another authorization decision. Name who may restore the agent, what must be reviewed and how the team confirms that the incident condition has ended. Preserve the original deactivation evidence when access returns. A useful kill-switch runbook explains the exact boundary it can stop, what remains outside that boundary and what the next operator must verify. It should not promise process termination or universal session cancellation without supporting evidence.
Frequently asked questions
Does Okta console deactivation end active sessions?
The Jamf account says current admin-console deactivation blocks new sessions. It separately describes gateway deactivation as revoking active tokens and ending sessions in flight, with the gateway kill switch landing in Q4. Preserve that distinction rather than assigning the gateway effect to the console.
Does Entra deactivation have the same documented effect?
Microsoft documents revocable agent identities, but its documentation does not establish identical active-token and in-flight-session effects for an Entra deactivation procedure. Request the procedure for the selected identity and resource, then test new access and existing access separately.
What should a kill-switch test retain?
Retain the agent identity, resource, stop action and observed outcomes for both new and established access. Add the relevant decision records and the restoration authority. State untested paths explicitly instead of treating a single blocked request as proof of universal termination.