Guides

AI agent kill switch: how to stop a run that goes wrong

CrowdStrike and DigitalOcean just shipped ways to stop misbehaving agents. Here's what a working kill switch needs and how a small team can set one up.

S

SelfAgentic Team

7 min read
A worn red push button below a metal plate reading MOTOR STOP on a scuffed grey panel.

On November 3, CrowdStrike added an AI agent kill switch to Falcon Guardian, letting an administrator stop an agent that has gone off course. Two days later DigitalOcean launched Cuttlefish, an iPhone app for approving, pausing and stopping agent runs from wherever you happen to be.

The question behind both launches applies to a three-person team too. If an agent misbehaves at 2 a.m., how fast can you stop it, and what does stopping involve? A working kill switch ends the run, removes the agent's access so it can't start again, and leaves a record of what it did. If you can do all of that in a couple of minutes, you have one. If any part needs a developer, you don't yet.

What is an AI agent kill switch?

An AI agent kill switch is any control that halts an agent mid-task and keeps it halted. The name comes from machinery, where a big red button cuts power whatever the machine is doing.

CrowdStrike's version is a useful reference because the press release spells out what "stop" means. An authorised admin can terminate the running agent (at the gateway first, with endpoint support described as coming soon), block it from relaunching on the targeted hosts, cut its enterprise network access, and log each of those actions to an audit trail. CEO George Kurtz's pitch: "When you deploy agents with a kill switch, you stay in control."

DigitalOcean's app works inside the agent session instead of on the network. According to the Cuttlefish launch announcement, you get a notification when a run needs approval, finishes or fails. You can approve or deny a tool call after seeing the full command, file path or diff, and you can redirect, pause or stop the run. Requests the app classes as high risk need Face ID or your passcode.

How to stop a rogue AI agent: controls to set up first

Most people picture a kill switch as a stop button, and stopping is the easy part. The harder questions are what the agent can still reach afterwards and what it already did. OWASP's GenAI Security Project put it bluntly in its Q3 2026 exploit roundup. About one incident it wrote, "The agent later stopped, but that did not undo the breach." After another, its advice was to "test a stop mechanism that revokes both credentials and network access."

You don't need enterprise security software to cover most of this. You need a few controls in place before anything goes wrong, because nobody wants to be reading documentation at 2 a.m.:

  1. A way to pause or stop a run from your phone. If your platform only lets you stop agents from a desktop dashboard, find that out now.
  2. Per-agent access you can revoke in one place. Each agent should hold its own grant to each tool, so cutting one agent off doesn't mean disconnecting the tool for everyone.
  3. Approval gates on outward-facing actions. Anything that posts publicly, emails a customer or changes ad spend should wait for a person by default. A queued action is easy to cancel. A sent email is gone.
  4. Hard limits on spend and volume. A token budget or a cap on actions per hour stops a looping agent even when nobody is awake to see it.
  5. A run record. Log every tool call and every token per run, so after a stop you can list exactly what to undo.

That list is close to how we built SelfAgentic. Connecting a tool and authorising a particular agent to use it are separate steps, and revoking a connection immediately removes every grant that depended on it. Publishing actions such as LinkedIn and Instagram posts or Meta ad changes need approval by default, and unattended runs wait in an Approvals inbox until someone decides. Our security page has the details, including what we don't claim yet.

Worked example: an overnight outreach run that goes wrong

Say a three-person B2B startup runs a workflow every night at 1 a.m. A research agent pulls 40 new leads into a Google Sheet, then a writer agent drafts a LinkedIn post and a first-touch email for each one.

At 1:20 a.m. the research agent reads a page with a broken data table and starts writing the same company into the sheet over and over. The writer agent picks up every new row and starts drafting.

With the five controls above, the night goes like this:

  1. The research agent hits its hourly action cap and the run halts. Nothing loops until morning.
  2. The writer's public posts and emails sit in the approval queue. None have gone out.
  3. At 7 a.m. the founder opens the overnight summary. It shows one failed run and 90 near-identical drafts waiting, and that one failure is the thing she reads first, ahead of the healthy totals.
  4. She rejects the drafts and revokes the research agent's access to the sheet while she checks the source page, so tonight's schedule can't repeat the mistake.
  5. The run record lists the rows the agent wrote, so cleaning the sheet takes about ten minutes.

Without those controls, the same night could end with 90 duplicate posts on the company page and a cleanup that eats most of the day.

What a kill switch won't fix

A kill switch responds to a problem. It only helps once someone, or some rule, notices. If nobody reads the alerts, the switch never gets pulled, so for a small team automatic limits do more work than any button.

It can't reverse actions that already reached the outside world. You can delete a published post, but people may have seen it, and you can't recall an email. This is the case for putting approval in front of outward actions and treating the stop as a backup.

The newest tools are also early. CrowdStrike describes endpoint termination as coming soon, and DigitalOcean's release says Managed Agents is in public preview, with preview services unsupported for production workloads. Plan around what ships today.

A kill switch also says nothing about why the agent went wrong. You still have to read the run and fix the cause before turning the schedule back on.

FAQ

What is the difference between pausing and killing an AI agent?

Pausing holds the run where it is so you can correct it and carry on; Cuttlefish, for example, lets you pause, send a correction and resume. Killing ends the run, and a full kill switch also blocks relaunch and removes the agent's access.

Do small teams need an AI agent kill switch?

Yes, if an agent runs unattended and acts in tools other people see. For a small team it's usually a mix of revocable per-agent access, approval gates and hard limits rather than a separate security product.

How do I revoke an AI agent's access quickly?

Give each agent its own grant per tool instead of a shared login, so you can remove one agent's access without disconnecting the tool. Revoking the underlying connection should cancel every grant that depends on it.

Can I stop an AI agent from my phone?

Some platforms support it now. DigitalOcean's Cuttlefish lets Managed Agents customers approve or deny tool calls and pause or stop runs from an iPhone. Before you rely on your own platform overnight, check that failure and approval alerts actually reach your phone.

Test your kill switch before you need it

Before your next unattended run, rehearse the stop once on purpose. Stop an agent mid-run, revoke its access to one tool, and check that the run record shows what it did. If any step takes more than a couple of minutes, fix that before you schedule more work.

If you want a team of agents with per-agent permissions, approval gates and hard token budgets in place from the first run, you can start building with SelfAgentic.

Keep reading

Ready to build your AI team?

See how a team of agents handles real work, and checks with you before anything goes out.

Free to start. No credit card required.