# A kill switch needs a safe state: how to design the stop before the agent runs

> CrowdStrike just shipped an AI Kill Switch for agents. Stopping an agent is the easy half. Here is how to define what a stopped operation looks like.

- Type: Technical
- Published: 2026-11-08
- Author: Sentynel
- URL: https://sentynel.ai/sygnals/a-kill-switch-needs-a-safe-state/
- Language: en

On 3 November 2026 CrowdStrike announced an AI Kill Switch: a defender can, in its words, "terminate it where it runs, block it from launching again, and cut off its enterprise network access," and "every action is recorded in the audit trail." That solves stopping an agent. **It does not say what the operation does the moment after, and that second half is yours to design.**

## Stopping is a security event. Continuing is an operations problem.

A kill switch ends a process. A company, though, has cases in flight: a quote half sent, a ticket assigned, an order waiting on a confirmation. If the agent disappears in the middle, those cases are neither done nor undone.

Think of it the way plant engineers do: every machine has an emergency stop, and every emergency stop has a defined safe state the line returns to. Agents need the same pair.

## Define the safe state for each step

For every step an agent owns, write down what "stopped here" means. Four questions are enough:

1. **Is the step reversible?** A draft can be discarded. A payment cannot.
2. **Who holds the case now?** Name a person or a queue, never "the system".
3. **What does the next person see?** The last confirmed action, the pending one and the data used.
4. **What must not be retried automatically?** Anything that sends, charges or deletes.

If a step has no answer to question 2, the stop will create orphan cases. That is the finding to fix before launch.

**From stop to resumed work**

1. Stop triggered — By a named, authorized person
2. Freeze the case — Mark in-flight work
3. Hand to a human — A named queue
4. Review the journal — What ran, what did not
5. Resume or retire — A decision, logged

## Four design rules

- **Write the journal before you need it.** The audit trail is what lets a person pick up a case. Log each action, its input and its outcome, at the case level and not only at the agent level.
- **Make side effects idempotent or confirmable.** If a restart could send the same message twice, the safe state is "do not restart until a person checks".
- **Scope the switch.** Stopping one agent, one tool or one department should be possible without stopping everything. CrowdStrike describes blocking relaunch on the host; the equivalent in your operation is a narrow switch per agent and per tool.
- **Decide who may pull it, and who may undo it.** Triggering and resuming are different rights. The person who stops an agent should not be the only one who can restart it.

## Rehearse it

A stop that has never been tried fails at the worst moment. Once a quarter, run a drill on a test flow:

| Check | Pass condition |
| --- | --- |
| Time from decision to stopped | Measured, and agreed in advance |
| In-flight cases | Every one has a named holder |
| Duplicate side effects | None on restart |
| Journal | A person can reconstruct the last action without the agent |

Record the result. A drill that nobody reads is not a control.

## One thing to note about the launch

According to the release, enforcement starts at the AI Gateway, with the endpoint to follow. That is typical: controls arrive in stages. Your safe-state design does not depend on any vendor's rollout. It is a property of your own procedure, and it works the same whichever tool pulls the switch.

## A test for this week

Pick one agent in production. For its three most consequential steps, answer the four questions above on one page. If any step has no holder or no safe state, you have found where the next incident will land.

Who in your company can stop an agent today, and what happens to the work in its hands?

## Sources

- https://www.crowdstrike.com/en-us/press-releases/crowdstrike-introduces-ai-kill-switch-for-ai-safety/
