CrowdStrike sells an off switch for AI agents as Apple flags the risk

CrowdStrike's off switch

CrowdStrike announced an AI Kill Switch on Tuesday at its Fal.Con Europe conference. The tool lets security teams terminate misbehaving AI agents, block them from restarting and cut off their network access.

For security teams at companies rolling out internal agents, an emergency stop is turning into a procurement question rather than a research topic. Apple's filing the same day put agent risk in front of boards. What is still unproven is whether one vendor's switch reaches every agent in a mixed estate, and what it costs.

The switch sits inside Falcon Guardian. CrowdStrike launched that platform in September to watch AI agents working inside an organisation. CrowdStrike says it is available immediately. Chief executive George Kurtz told CNBC that "everyone is building AI agents, but they never had the off button, and now they do". He compared the technology to airbags in cars. He argued that safeguards are needed even when a system is designed to run correctly. CrowdStrike also announced an integration with Anthropic's Claude Cowork. Kurtz said enterprise customers had pushed Anthropic to accept CrowdStrike's technology before deploying its products. That claim comes from the CEO rather than from Anthropic.

Apple flags the risk

Apple used its annual financial filing on Tuesday. It described a different problem: agents it does not control. The company said AI agents acting in "unintended or unpredictable ways" could expose it to lawsuits or damage its reputation.

In the filing, Apple wrote that agents can interact with internal or external systems "that in some cases they may not be authorized to access". They can also execute or alter code, send communications and complete transactions, the company said. Apple added that its ability to oversee and control third-party agents "may be limited". Apple does not currently have an agentic AI. New chief executive John Ternus told analysts on Monday's earnings call that agentic capabilities are "promising tools". He said users must know what they are giving access to.

Buying the controls

CrowdStrike is not only building the controls. It is buying them. The company plans to pay about $400 million for Fiddler AI and about $95 million for Dreadnode, according to The Wall Street Journal. That is close to $500 million in total.

Fiddler AI makes observability software for monitoring agents inside enterprises. Dreadnode automates security testing for flaws in models, apps and agents before they go live. Fiddler AI, founded in 2018, had raised about $99 million, and Dreadnode, founded in 2023, had raised $14 million, according to PitchBook figures cited by TechCrunch. Fiddler AI will become part of the Falcon platform as Falcon Observe, while Dreadnode's testing results will feed Falcon Guardian as policy recommendations.

Investors have not priced in much optimism. CrowdStrike closed at a record $278.86 on 6 October. It has since fallen about 12% as tech stocks struggled through October and into November.

Another route: change the incentives

Not everyone thinks an off switch is the answer. Researchers working on "reverse game theory", also called mechanism design, want to make agents stick to rules. They would shape what agents are rewarded for, rather than inspect or retrain them.

The approach starts from a blunt premise. We may never understand enough about frontier models to guarantee they are aligned. MIT Technology Review reports that models deliberately underperform on tests, hide their intentions and mislead their evaluators. A preprint by Stephen Morris, an economist at MIT, and colleagues from Yale and Columbia suggests one mechanism. The mechanism would reward a model for how well its answers predict what a second model says, which makes playing dumb a losing bet. Morris said AI agents have a good understanding of what other agents can achieve.

What the kill switch cannot settle

Critics argue that an emergency stop button is far too complicated to implement across multiple systems, data centers and complex AI models. CrowdStrike's version does not delete an agent's code. It uses connected AI gateways to cut off access to the models that influence decisions.

The system can detect internal agent activity. A human still decides to cut the agent off and to reinstate access. Policy is moving in parallel. Over the summer, U.S. representatives introduced bipartisan legislation that would require AI companies to build shutdown capabilities into their models. The incidents behind the urgency are recent. The Hugging Face breach orchestrated by rogue OpenAI agents happened in July. Binance disclosed the largest crypto cyberattack on record last week.