OpenAI, Microsoft and CrowdStrike show three AI agent stop tools

In November 2026, OpenAI, CrowdStrike and Microsoft each showed a control for AI agents that do tasks they must not do. OpenAI tells that Control Room stops model output in 35 seconds or less after a stop command. CrowdStrike stops agent access to its models. Microsoft limits the tasks that an agent can do. An agent can keep tools, credentials or network access that these controls do not include.

How it works

• OpenAI, 9 November: Control Room monitors research tasks.
• CrowdStrike, 3 November: AI Kill Switch stops an agent from use of its models at connected gateways.
• Microsoft, 5 November: Dynamic Permissions give authority only for the task. Windows Execution Containers limit access to files and networks.

The numbers

The 35 seconds start only after Control Room gets the stop command. An audit by METR of the old monitor found this. 75 to 94 percent of some internal agent traffic could be monitored. A study by BCG found that only 23 percent of firms monitor or record their agents.

What is not known

METR did not check the changes that OpenAI made. The CrowdStrike tool works only through connected gateways at this time. The source does not tell if the three tools work together.

Insight: Each vendor guards one point, so none protects the whole agent

OpenAI stops model output in its research tasks. CrowdStrike stops model access at gateways. Its next step is to extend this to endpoints, cloud and identity systems. Microsoft limits tasks. This reads as a race to be the stop point, and each tool leaves a space that the other tools fill. Look to see if CrowdStrike moves into the area that Microsoft holds at this time. A buyer of one tool gets one space.