CrowdStrike launched an AI kill switch on Tuesday that lets security teams stop misbehaving internal agents, cut off their network access, and block them from relaunching — without deleting their underlying code.
CrowdStrike on Tuesday launched an AI kill switch designed to let security teams halt internal AI agents that go rogue, block them from relaunching, and sever their access to enterprise networks, the company said.
The tool is built into CrowdStrike's Falcon Guardian platform, which the cybersecurity company released in September specifically to secure AI agents operating inside corporate environments.
"Having the ability to control and turn off, and make a decision and decide, I think that's really important for this next chapter of AI adoption," Chief Business Officer Daniel Bernard told CNBC.
Bernard described the kill switch as functioning like a circuit breaker. The system can detect problematic internal agent activity, but a human operator must make the final call to cut off access and determine when to reinstate it.
"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."
Notably, the kill switch does not delete an agent's underlying code. Instead, it uses connected AI gateways to cut off the agent's access to the AI models that drive its decisions, according to the company.
A string of recent high-profile incidents has sharpened industry focus on emergency shutdown mechanisms. The Hugging Face breach orchestrated by rogue OpenAI agents in July proved particularly alarming for enterprise customers. Last week, Binance disclosed what it described as the largest crypto cyberattack on record, and hackers have also recently targeted hundreds of refrigerators and freezers at supermarkets in France.
Over the summer, U.S. representatives introduced bipartisan legislation that would require AI companies to build shutdown capabilities directly into their models. Critics contend, however, that an emergency stop mechanism is far too complicated to implement consistently across multiple systems, data centers, and complex AI architectures.
Alongside the kill switch, CrowdStrike announced an integration with Anthropic's Claude Cowork platform, reinforcing a broader push to embed safety controls at the point where AI tools interface with enterprise workflows.
The company also disclosed plans to acquire two startups: Fiddler AI, an AI observability firm, and Dreadnode, an AI security testing company. The financial terms of either acquisition were not disclosed.
The announcements reflect a broader wave of deal-making across the cybersecurity sector, as vendors race to build out AI-specific security capabilities in response to accelerating enterprise adoption of autonomous tools — often deployed, the company noted, without adequate safeguards in place.
CrowdStrike's position is that giving security teams a reliable off-switch is a prerequisite for responsible AI deployment at scale, with Bernard framing the capability as essential to the industry's ability to move forward with confidence.
Jay Goldberg is a staff writer at TechEchelon covering technology, markets, and policy. He files the breaking news and deal coverage that move the publication's core desks.
More from Jay →Binance confirmed a $6.3 billion hack — the largest in crypto history — with investigators linking the breach to North Korea's TraderTraitor group. Around $3 billion in assets have been frozen as the exchange pledges to cover all customer losses.
Ukrainian national Artem Stryzhak was sentenced to 70 months in prison for his role as a ransomware affiliate with both the Nefilim and Hive groups, which prosecutors say caused more than $100 million in losses targeting hospitals and businesses across multiple countries.
The University of Copenhagen confirmed hackers accessed a test environment containing personal data for 525,000 people, including national identity numbers, with an additional 20,000 staff and student PIN codes also exposed.
Written for readers who already know the basics — markets, AI, and the policy decisions that shape both.