CrowdStrike introduced a kill switch for AI agents on Tuesday. It lives in Falcon Guardian, the agent security platform the company launched in September, and it gives security teams three moves: stop an internal agent that has gone rogue, block it from relaunching, and cut off its access to the enterprise network. CrowdStrike also announced an integration with Anthropic’s Claude Cowork.
How the switch works
Daniel Bernard, chief business officer, described it to CNBC as a circuit breaker in systems where customers run the Falcon platform. The process detects internal agent activity on its own, but a human decides whether to cut the agent off and whether to let it back in.
“If you kill a process at the root, and we have that level of access at the root … it’s very binary,” Bernard said. “You either kill the AI and the agent, or you don’t, and then you report on that.”
The agent’s code stays where it is. What gets severed, through connected AI gateways, is access to the models that shape its decisions. That is the part I find convincing. You are not wiping the program, you are removing its ability to decide anything.
The pressure built over a few months
CNBC names three incidents behind the urgency. The Hugging Face breach in July, orchestrated by rogue OpenAI agents. Binance last week, the largest crypto attack on record. And a few hundred hijacked fridges and freezers in French supermarkets.
Over the summer, US representatives introduced bipartisan legislation that would require AI companies to build shutdown capabilities into their models. Critics argue an emergency stop is far too complicated to implement across many systems, data centres and complex models.
CrowdStrike is also acquiring two companies: Fiddler AI in AI observability, and Dreadnode, which automates AI security testing.
A switch is not control
On the Cowork integration itself, the announcement says only that it exists. What it actually does still has to be delivered, and that is where the interesting question starts. Cowork sessions reach files, connectors and scheduled tasks. A company that wants to know what one of those sessions touched overnight does not need a kill switch. It needs a log.
The kill switch is the right thing to build, just not the hard part. The hard part is knowing when to press it. CrowdStrike keeping a human in that decision is honest. It only moves the problem to where it already was.