At Fal.Con Europe in Barcelona on Tuesday, CrowdStrike made a move that would have looked absurd two years ago: it acquired two AI security startups and launched a literal kill switch for AI agents, all in one day. The message is blunt. Enterprises are scared of the agents they're deploying, and CrowdStrike intends to sell them the entire safety stack.
The acquisitions bracket the agent lifecycle. Dreadnode extends the Falcon platform into pre-deployment testing, running more than 100 known AI attack strategies — jailbreaks, prompt injection, multi-turn and agent-specific attacks — against the models, agent reasoning loops, applications, and MCP servers that power enterprise AI, according to CrowdStrike. Findings get ranked by severity and mapped to OWASP, MITRE ATLAS, and NIST frameworks so teams can make a go/no-go decision before an agent ever reaches production. The Dreadnode team joins CrowdStrike's Cyber Superintelligence Lab, and the deal — reported at about $95 million — is expected to close in Q4 FY27, paid predominantly in cash with some vesting stock.
Fiddler AI, announced in a separate Falcon Guardian release the same day, covers the other side of the lifecycle: observability — watching agents after they run. And then the product launch: the AI Kill Switch, available today, lets an authorized administrator terminate an unauthorized or misbehaving agent at the Falcon Gateway, block it from relaunching on a host or group, and strip its enterprise network access. Endpoint termination is planned next. The response stays human-triggered, with every step recorded in an audit trail, and CEO George Kurtz said the feature is intended to let companies "deploy agents while retaining control when their behavior changes," as ad-hoc-news reported.
The kill switch is the honest one. A test suite and a monitoring dashboard are table stakes; an off-switch admits the premise. "Enterprises are deploying AI faster than they can prove it is secure," Kurtz said in the Dreadnode announcement. "A model can be jailbroken, an agent can be manipulated, and most organizations have no way to find those flaws until the AI is already in production."
That premise has been the entire arc of this month. Anthropic cut its agents' internet access after admitting it couldn't watch them; OpenAI tapped the startup that breached Anthropic in 12 hours to run its own security sprint, announced this morning; the GDPX benchmark found the best agent on Earth scores 12% on enterprise work, with finance and legal at zero. CrowdStrike didn't invent the fear. It just packaged the answer — before deployment, after deployment, and at the point of execution: off.
What to watch: whether the kill switch reaches the endpoint layer, where CrowdStrike's sensor actually lives — and whether "human-triggered" scales when the agent fleet does. The switch only works if someone is watching.
— via CrowdStrike