CrowdStrike announced an AI Kill Switch on November 3 at its Fal.Con Europe event and says it is available now. It is part of Falcon Guardian, CrowdStrike’s product for finding AI agents in a company, setting rules for what they may do and detecting attacks against them. The switch is for one situation: an agent has gone off course and needs to stop now.
What it does today
- Cuts off the model. An authorised administrator triggers the switch in Falcon Guardian. Connected AI gateways then block the agent’s access to the models it uses to decide and act. This works with CrowdStrike’s Falcon AI Gateway or a third-party gateway.
- Blocks a restart. A Falcon Guardian policy stops the agent from launching again on a chosen host or host group.
- Removes network access. The press release says the agent loses its access to the company network.
- Leaves a record. Every action is written to the audit trail.
Falcon AI Gateway is included with Falcon Guardian at no extra cost. CrowdStrike does not list a separate price for the kill switch.
What is still coming
CrowdStrike lists several controls as “coming soon”: ending agent processes on the endpoint, containing agents in the cloud, revoking an agent’s permissions through identity security, stopping SaaS agents, ending browser sessions, and triggering the switch automatically from its SOAR product. The company’s own release asks customers to base purchases on features that are available today.
What it means for teams using agents
As coding and work agents get access to repositories, tickets and internal tools, security teams need a way to stop one quickly without shutting everything down. Today the control point is model access through a gateway, so it only covers agents whose model traffic passes through a connected gateway. If your agents call models directly, plan for that gap. Whatever product you use, it helps to keep a list of which agents run where and which credentials each one holds.
A simple drill is worth running before you need it: pick one test agent, decide who is allowed to stop it, trigger the stop, and check that it cannot restart or reach internal systems afterwards. The same drill shows whether your logs record who acted and when.
Related on ToolAI
For tools that put agents to work, see Cursor’s Agent SDK and Bitbucket Agentic Pipelines. For another side of AI in security, read how AI decoy bots waste scammers’ time.
Sources: CrowdStrike press release · CrowdStrike blog