An AI agent kill switch is the button, command, or admin control that stops an agent mid-task. We checked 10 real agents and ranked each one VERIFIED (the vendor publicly documented it), CLAIMED (the vendor says so, but we found no primary proof), or UNKNOWN (no public information found). Every entry cites its source.
Key figure: 63% of organizations in IBM's 2025 breach study had no AI governance policy, and 97% of AI-breach victims had no proper AI access controls (IBM, Cost of a Data Breach Report 2025).
VERIFIED The vendor publicly documented a kill switch. Source link cited.
CLAIMED The vendor or credible reporting says a stop control exists, but no primary documentation was found. Treat as unproven.
UNKNOWN No public information about a kill switch. Absence of evidence is not evidence of absence — it means we could not verify one.
Methodology: we searched vendor documentation, changelogs, and product pages on October 17, 2026. We never invent a claim — if it is not documented, it is UNKNOWN.
Showing 10 of 10 agents.
| Agent | Vendor | Kill-switch status | Evidence |
|---|---|---|---|
| Operator | OpenAI | CLAIMED | OpenAI says a monitor model "can pause the task if something seems off", plus takeover mode and user confirmations before significant actions. Controls described by the vendor, but no standalone kill-switch documentation found. Source: OpenAI, "Introducing Operator" |
| Copilot Studio agents | Microsoft | VERIFIED | Microsoft Learn documents governance controls: admins can Block an agent org-wide from the agent registry in the Microsoft 365 admin center. Source: Microsoft Learn |
| Claude Agent SDK | Anthropic | VERIFIED | The official SDK docs include client.interrupt(), which stops the current task programmatically — a documented kill switch for agents built on the SDK. Source: Anthropic SDK docs |
| Claude Code (CLI) | Anthropic | CLAIMED | Widely documented that Esc interrupts Claude mid-turn and Ctrl+C interrupts a running operation. This comes from community documentation, not Anthropic's primary docs, so it stays CLAIMED. Source: shortcut reference |
| Cloud Agents | Cursor | CLAIMED | Cloud agents run on Cursor's own machines and can be stopped or cancelled from the agents dashboard. Documented in Cursor's own docs as summarized by third-party analysis; we did not find a standalone vendor kill-switch page. Source: analysis of Cursor docs |
| Devin | Cognition | VERIFIED | Cognition's own product update documents a "stop" button at the top right corner of the chat to end sessions manually. Source: Cognition blog |
| Replit Agent | Replit | CLAIMED | Replit Agent offers a pause/stop button mid-task plus checkpoint rollback. Replit's own docs confirm checkpoints and paid-action confirmations; the stop control is documented by third-party guides rather than a primary vendor page. Source: guide · Replit docs |
| Manus | Manus | CLAIMED | Reporting describes a "Manus's Computer" window where humans can watch and interrupt tasks mid-execution. No primary vendor documentation of the interrupt control was found. Source: TechWafer |
| Grok Build | xAI | VERIFIED | xAI's own Grok Build changelog documents [stop] labels on killable rows and /workflow commands that pause and stop background runs. Source: xAI changelog |
| Gemini Agent | UNKNOWN | No public documentation of a user-facing kill switch for Google's Gemini Agent was found. (Google shut down Project Mariner on May 4, 2026 and folded it into Gemini — a vendor-level shutdown, not a user kill switch.) |
A documented kill switch is one of the few controls an underwriter can actually verify. It is also the difference between a contained incident and an autonomous runaway — which is why our prompt injection insurance calculator treats missing human controls as a top risk factor, grounded in real 2026 carrier signals: Armilla AI raised its standalone AI liability limits to $25M in January 2026, while ISO's generative AI exclusions (CG 40 47, CG 40 48, CG 35 08) took effect on commercial general liability policies on January 1, 2026.
Because agents act at machine speed and fail at machine scale. A runaway agent can send thousands of emails, spend real money, or corrupt data before a human notices — the kill switch is the difference between an incident and a catastrophe. This registry catalogs real agents’ stop mechanisms so you can verify yours exists, works, and is reachable in under a minute.
An AI agent kill switch is a control that lets a human immediately stop an agent mid-task — a stop button, an interrupt command, or an admin block. It is the last line of defense when an agent goes off-script, gets prompt-injected, or starts taking actions you never approved.
It depends on the agent. Local coding agents usually stop with Ctrl+C or Esc. Cloud agents are stopped from their dashboard — Devin documents a stop button in its chat UI, and Cursor's cloud agents are cancelled from the agents dashboard. Enterprise platforms like Microsoft Copilot Studio let admins block an agent org-wide from the agent registry.
In this registry, Devin (stop button documented on Cognition's own blog), Microsoft Copilot Studio agents (admin Block action documented on Microsoft Learn), xAI's Grok Build (stop labels documented in xAI's own changelog), and Anthropic's Claude Agent SDK (client.interrupt() in the official SDK docs) carry VERIFIED status. Others are CLAIMED or UNKNOWN — check each entry's cited source.
Insurers price agent risk on controls. Armilla AI's standalone AI liability policy — limits raised to $25M in January 2026 — is built around how AI systems actually fail, and standard commercial general liability policies now commonly exclude generative AI via ISO endorsements CG 40 47, CG 40 48 and CG 35 08. A documented, working kill switch is one of the clearest controls an underwriter can point to.