Five switches, actually. Four scope a stop to exactly what has gone wrong: a user, a device, a provider, or a tool. The fifth halts AI across every enrolled device and provider in one move. All immediate, all reversible, all recorded.
Sometimes the problem is one agent: it's autonomous, it's mid-run, and it's about to take a step you never approved. Stopping it means stopping the user or device it runs under, and that lands before its next action. Scope a stop by user, device, provider, or tool, each one available inline in the audit log where you spotted the problem. And when the problem is wider, a compromised tool, an injected instruction, or a decision to pause AI while you investigate, a fifth switch halts everything. None of them wait on a vendor.
A stop you'd hesitate to use is no stop at all. These are built to be pulled without ceremony and cleared without cleanup.
An autonomous agent doesn't get to finish what you didn't approve. Scope the stop to the user, device, provider, or tool involved, and the next action is refused mid-run, so the step it was about to take is the step that never happens.
One move halts every connection through the checkpoint, on every enrolled device and for every provider, with no parsing in the way and no vendor's cooperation needed. The tools nobody registered go quiet with everything else.
The activation itself lands in the signed record, and so does the moment you bring things back. Later, you can show exactly when AI stopped, who stopped it, and when it resumed.
Watch the whole arc: traffic flowing, containment landing, the fleet coming back. Each next connection is refused at the checkpoint the moment containment lands, and every transition seals its own entry in the chain.
The goal is not to stop AI. It is the opposite, to let institutions finally say yes.