A kill switch that works at the AI gateway can only revoke one of the three things a running agent holds — and it is the only one of the three that was never yours to lose. CrowdStrike shipped exactly that on 3 November 2026, which makes this a good week to find out how long your own credential revocation takes, because nobody is going to ship that switch for you.
What shipped, and where it enforces
CrowdStrike announced the AI Kill Switch on 3 November 2026 at Fal.Con Europe in Barcelona, available immediately in Falcon Guardian. The company frames it as giving security teams "the ultimate human control over AI agents", and the mechanism is specific: a team activates the switch in Falcon Guardian, and it directs connected AI gateways to block access to the AI models that power agents' decisions and actions. Reporting on the launch adds that it can also stop an agent from relaunching on a host group and restrict its access to enterprise networks, and CrowdStrike describes it as a human-operated failsafe rather than an automated response — though Charlotte Agentic SOAR is positioned to orchestrate automated kill-switch workflows triggered by AI-specific detections, which the company lists as coming rather than shipped.
The enforcement point is the whole story. An AI gateway sits in front of model traffic, which makes it the natural place to switch inference off: one policy change, every agent behind it, no agent cooperation required. It is a genuinely good design. It is also a design whose reach is defined by what flows through the gateway — and an agent's capacity to cause harm is not mostly made of model calls.
An agent holds three grants, and they revoke independently
Decompose what a running agent actually has. There are three grants, they come from three different places, and stopping one does nothing to the others.
- Inference. The ability to produce a new decision — a token, a plan, a tool call. Granted by a model provider, mediated by your gateway, and the thing a gateway-enforced switch revokes.
- Authority. The credentials the agent presents to act: OAuth tokens, service-account keys, session cookies, signed requests. Granted by your identity system and your credential broker, and completely untouched by anything that happens at a model gateway.
- Effects. The writes that already landed — the row inserted, the email sent, the branch pushed, the payment authorised. Granted by nobody and revoked by nobody; effects are undone only by a compensating action you wrote in advance.
Gartner's Apeksha Kaushik called the product an emergency brake rather than a rewind button, which is the same observation in one sentence. Forrester's Pollard made the adjacent point: kill switches are a useful control for responding to agentic incidents, not for preventing them. Both are endorsements of the thing as what it is, and both are warnings about the reading the name invites.
What keeps running after inference stops
"Block the model" sounds total because we picture an agent as a loop whose every step begins with a model call. Stop the call, stop the loop. That picture is right for a chat-shaped agent and wrong for most of what is actually deployed, because production agent systems are full of machinery that runs without asking a model anything.
Work already in flight
A tool call dispatched one millisecond before the switch fired completes normally. If that tool was a batch job, a migration, an outbound send or a workflow trigger, the model's silence is irrelevant — the effect is downstream of a decision already made. An agent whose last act was to enqueue two hundred messages has, from the gateway's point of view, stopped. From the mailbox's point of view it has not.
Scheduled and triggered entry points
Anything that wakes an agent on a timer or a webhook keeps firing. The runs will fail at the model call, which is the intended outcome, but the retry logic around them usually is not idempotent in the way you would want during an incident — the idempotency and retries problem becomes an availability problem at exactly the moment you were trying to reduce activity. A switch that blocks inference turns a misbehaving agent into a retry storm unless something also disables the trigger, which is a separate control in a separate system.
Code the model wrote earlier
This is the one most likely to be missed. An agent that writes and schedules a script, installs a hook, creates a CI job or registers a cron has converted model output into ordinary automation that no longer needs a model. Agent artifacts on the endpoint covers the inventory problem; the point here is that cutting inference does not cut the artifacts, and the artifacts are exactly what an agent that was pursuing a wrong goal for a while will have left behind.
Blocking inference bounds new decisions. It does not bound the blast radius of decisions already taken — that is the subject of blast radius and it is set before the incident, by what the agent was allowed to reach.
Gateway enforcement is a scope statement
The second consequence of the enforcement point is about coverage rather than timing. A gateway can only block traffic that traverses it, which partitions your agent estate into two sets whose boundary has nothing to do with risk.
In scope: agents your teams built, pointed at your gateway, using your keys. That is the population a platform team can see, inventory and already mostly governs. Out of scope: every agent embedded in a SaaS product you bought, calling its vendor's model over its vendor's network, acting inside your tenant with credentials you granted at install time. The AI features in your CRM, your ticketing system and your code host are agents by any useful definition, they hold real authority in real systems of record, and no switch in your security console reaches their inference path.
That asymmetry is uncomfortable because it inverts the usual intuition. The agents you can kill are the ones you understand best and whose permissions you scoped yourself. The agents you cannot kill arrived with a vendor's default permission set and a consent screen — the situation in third-party agents logging in as your users. For those, the only revocation you hold is the second grant: withdraw the token, kill the session, remove the app. Which is the same conclusion from a different direction — the control that generalises across your whole estate is credential revocation, not inference revocation.
An unmeasured revocation is not a control
MLQ.ai's write-up made the observation that matters most for anyone putting this in a runbook: there is no published third-party test of how quickly the switch acts. That is not a criticism of the product so much as a description of the category. Every revocation in an agent system has a latency, almost nobody measures it, and the number is routinely much larger than people assume.
Inference revocation at a gateway is bounded by policy propagation and by whatever connection pooling sits between the agent and the gateway. Credential revocation is worse, and in a predictable way: a bearer token already issued remains valid until it expires unless the resource server checks a revocation list on every request, which most do not. That is the whole argument of credential lifetime — the honest revocation window for an OAuth access token is its remaining TTL, not the moment you clicked revoke. An agent cut off from its model at 14:02 with a token good until 14:47 is an agent you have stopped thinking and have not stopped acting.
So the thing to produce this quarter is not a longer control list. It is three numbers, measured on your own stack in a game day, with a stopwatch:
| Grant | What you build | How you test it | The number you need |
|---|---|---|---|
| Inference | Gateway deny rule by agent identity | Fire it mid-run against a live agent | Seconds from click to first blocked call |
| Authority | Broker revocation plus short TTLs | Revoke, then replay a real request | Worst-case residual validity, not mean |
| Effects | Compensating action per tool | Undo a real write in staging | Share of writes with a tested undo |
The third row is the one most organisations cannot fill in at all, and it is the row that decides what an incident costs. Repairing agent side effects is the engineering; decision receipts and audit is how you know which writes to repair, because an undo you cannot scope is an undo you will not run.
What to change this week
Buy the switch if you have the platform — a human-operated brake on inference is strictly better than nothing, and the gateway is the right place for it. Then treat its arrival as a prompt to close the two gaps it reveals.
- Measure credential revocation before you trust inference revocation. Pick one production agent, revoke its token, and time how long a replayed request keeps succeeding. If the answer is "until the token expires", shorten the TTL; that is the cheapest change in this post and the one that most improves a real incident.
- Pair every kill path with a trigger disable. A blocked model plus a live scheduler is a retry storm. Put the trigger kill in the same runbook step, not a different system's console.
- Inventory the agents the switch cannot reach. Vendor-embedded agents calling vendor models will not appear in your gateway logs. For those the kill switch is the app-consent screen, and the revocation drill is different. Detecting agent compromise covers the signals you have left when the inference path is not yours.
- Write the compensating action with the tool, not after the incident. Any tool that mutates a system of record ships with a tested undo or it ships behind a human gate. That rule costs you velocity once and saves you the row in the table you otherwise cannot fill.
Alongside the announcement, CrowdStrike agreed to buy two AI security startups for a combined $495 million according to the Wall Street Journal, one of them Dreadnode, which builds tooling to test AI systems for security vulnerabilities — and Microsoft, Palo Alto Networks and Cisco are all reported to be building comparable controls. The category is arriving fast, and that is good. Just read every product in it by its enforcement point, and ask which of the three grants it actually takes back.
FAQ
Does CrowdStrike's AI Kill Switch stop an agent completely?
It stops the agent from getting new model output, by directing connected AI gateways to block access to the models behind its decisions, and reporting indicates it can also prevent relaunch on a host group and restrict enterprise network access. It does not invalidate credentials the agent already holds, and it does not reverse writes already committed.
Why is credential revocation harder than blocking the model?
A bearer token that has already been issued stays valid until it expires unless the resource server checks a revocation list on every request, which most do not. So the practical revocation window is the token's remaining lifetime, which is why short TTLs matter more than a revoke button.
Does a kill switch help with agents inside SaaS products we buy?
Generally no. A gateway can only block traffic that passes through it, and a vendor-embedded agent calls its vendor's model over its vendor's network. Your lever there is withdrawing the app's authorisation or its tokens.
How fast does the switch act?
No public third-party test of its time-to-effect has been published, which is normal for the category. Measure it on your own stack in a game day rather than assuming it is instant, and record the worst case rather than the average.
If the switch cannot undo anything, is it worth having?
Yes — bounding new decisions during an incident is genuinely valuable, and an auditable human brake is better than persuading an agent to stop. The error is treating it as containment. Gartner's framing, an emergency brake rather than a rewind button, is the right mental model.
Further reading
On this wiki:
- Kill switches — the control-design version of this argument, independent of any vendor.
- AI gateways — what sits at the enforcement point and what else it can do there.
- Credential lifetime — why revocation latency is mostly TTL.
- Repairing agent side effects — the third grant, and the only way to take it back.
- Blast radius — set before the incident, not during it.
Sources:
- CrowdStrike, "CrowdStrike Introduces the AI Kill Switch for AI Safety"
- CrowdStrike blog, "CrowdStrike Introduces AI Kill Switch"
- CSO Online, "CrowdStrike launches AI kill switch to shut down rogue agents"
- MLQ.ai, "CrowdStrike launches AI kill switch, but full agent containment is still ahead"
- Forbes, "CrowdStrike Looks To An AI Kill Switch To Beat Rogue Agents"