Insights

AI Governance

AI agent kill switches are a product now, so write the runbook that uses one

CrowdStrike's Nov. 3 AI Kill Switch stops an agent, blocks it from relaunching and cuts its network access, with an authorized administrator pulling the trigger. The control is the easy part. HealthTech vendors still need a one-page runbook for who stops which agent, for which customer, and how they prove it stopped.

November 10, 2026 · 6 min read

AI agent kill switches are a product now, so write the runbook that uses one

Stopping an AI agent now has a product category. On Nov. 3, CrowdStrike announced an AI Kill Switch in Falcon Guardian that terminates a running agent, blocks it from launching again on a host or host group, and cuts off its enterprise network access. An authorized administrator triggers it, and each action is recorded in an audit trail.

The same day, CyberScoop reported that CrowdStrike signed agreements to acquire Fiddler AI, which does agent observability, and Dreadnode, which does automated security testing of models and agents. Both deals are expected to close in CrowdStrike's fourth fiscal quarter, pending standard approvals.

This is not a product review, and I am not recommending a tool. The point is narrower. When a large security vendor ships "stop the agent" as a feature, health systems will start asking their own vendors the same question: if your AI misbehaves in our environment, who stops it, how fast, and how do we know it stopped? A product gives you a mechanism. It does not give you the answers.

What the announcement actually describes

The press release lists four parts, and they are a useful checklist even if you never buy the product:

  • Terminate the running agent. The release says this happens "first at the Gateway, and soon on the endpoint," so the endpoint enforcement is described as coming, not shipped.
  • Block it from running again, by policy, on the targeted host or host group.
  • Cut its network access.
  • Keep a human in control, with every action logged.

CEO George Kurtz, writing on X and quoted by CyberScoop, said the switch "lives outside the model, at runtime, where an agent can't argue with it, rewrite it, or wait it out." That design principle is the part worth keeping. A stop control that the agent can reach is not a stop control. I made a related argument about agents and their own monitors: anything the agent can write to, it can eventually route around.

Why vendors need their own version

The CrowdStrike control is built for an enterprise stopping agents on its own endpoints and network. A HealthTech vendor has a different problem. Your agent runs in your infrastructure, on behalf of many health systems, often touching each one's data through integrations. When one customer calls and says "stop it," they mean stop it for them, now, without taking down everyone else.

That request has three hard parts: scope, authority and proof. A kill switch product helps with the mechanism inside an environment you control. It does not tell you whether your architecture can isolate one tenant, who on your side may act without a vice president's approval at 2 a.m., or what evidence you send the customer afterward.

Health systems already think this way. Their governance teams want to know how to turn a tool off, and their incident plans assume they can. Last week I wrote about OpenAI's sandbox incident, where detection was fast and the stop was not, because nobody had decided in advance what the default should be. The runbook is how you make that decision ahead of time.

The one-page stop-the-agent runbook

Write one page per agent or AI feature. Keep it short enough that an on-call engineer can follow it without reading anything else.

  1. Trigger authority. Name the roles that may stop the agent without further approval, and the roles that may resume it. Stopping should need fewer approvals than resuming. Include who on the customer side can request a stop, and how you verify the request is genuine.
  2. Default rule. State what happens when a top-severity alert fires and nobody has ruled it out within a set window. My recommendation is pause by default. Pausing is reversible. A long debate is not.
  3. Scope levels. List each level you can actually stop at: one customer's tenant, one integration, one tool the agent can call, one model version, or the whole feature. For each, write the exact mechanism, such as a feature flag, a revoked credential, or a blocked network route. If you can only stop at the whole-feature level, write that down. It is an honest answer and a design gap.
  4. Containment beyond the process. Stopping the process is one step. Revoke or rotate the agent's credentials and tokens for the affected scope, block its outbound network paths, and cancel queued jobs so they do not run when something restarts. The CrowdStrike release pairs termination with a relaunch block and a network cut for the same reason.
  5. Verification. Define how you prove it stopped: no new actions in the agent's action log after a timestamp, credential use attempts failing, no traffic from the agent to the customer's endpoints. Name who checks and record the time.
  6. Customer notice. Draft the message now. It should say when the agent was stopped, at what scope, what it did in the window before the stop, and when the customer will hear next.
  7. Resume criteria. List what must be true before the agent runs again for that customer, and who signs off.

Put a version number and a date on the page. Review it when the agent gains a new tool, a new integration or a new customer type.

How to know the runbook works

A runbook nobody has run is a guess. Test three things on a schedule.

First, the mechanism. In a staging environment, stop the agent at each scope level you listed and time it. If tenant-level stop takes longer than the default window in step 2, you have found your first fix.

Second, the people. Run a tabletop where the alert arrives outside business hours and the person with stop authority is unreachable. See whether the next person on the list knows they have the authority.

Third, the evidence. After each test, check that the action log, the credential revocation and the customer notice draft all carry timestamps that line up. If you cannot build a clean timeline from your own test, you will not build one during an incident.

Where products fit

Products like the one CrowdStrike announced will make the mechanism part easier for some teams, and for agents running on managed endpoints that matters. Fiddler and Dreadnode, if those deals close as described, push the same vendor toward agent observability and pre-deployment testing, which are the inputs a stop decision depends on.

None of that writes step 1 or step 3 for you. Tool choice comes after you know which scopes you must be able to stop, who decides, and what you will tell the customer. If a health system asked you today to stop your agent for their tenant only, could you name the mechanism and the person in one sentence?