CrowdStrike's AI Kill Switch puts an off button on enterprise agents

A kill switch inside Falcon Guardian

CrowdStrike launched the AI Kill Switch on Tuesday at its Fal.Con Europe conference. It is available immediately. The tool sits inside Falcon Guardian, the agent security platform CrowdStrike launched in September.

For buyers, the pitch is not another dashboard that watches agents. It is a human-controlled off switch for agents already running inside corporate networks. That puts pressure on rivals selling agent observability to show enforcement at runtime, not discovery alone. It is still unclear how the switch behaves against third-party agents, and against systems built to survive a shutdown.

CrowdStrike says the switch lets security teams terminate a misbehaving internal agent. It also stops the agent from relaunching and cuts off its access to the enterprise network. A person decides when to cut an agent off and when to reinstate it, according to the company. The switch does not delete the agent's code. It uses connected AI gateways to cut access to the models that shape the agent's decisions, CrowdStrike said.

"Everyone is building AI agents, but they never had the off button, and now they do," CEO George Kurtz told CNBC's Jim Cramer. Chief business officer Daniel Bernard called the mechanism a circuit breaker. "You either kill the AI and the agent, or you don't, and then you report on that," he told CNBC.

CrowdStrike also announced an integration with Anthropic's Claude Cowork. Kurtz told CNBC that customers "would not roll out AI without CrowdStrike".

Buying the observability layer

The launch came with two acquisitions. CrowdStrike is not buying alone: it and its rivals have been scooping up cybersecurity startups as demand for AI security tools grows.

Fiddler AI makes observability software for monitoring AI agents inside enterprises. Dreadnode automates security testing aimed at finding flaws in AI models, apps and agents before they go live. Fiddler AI will be folded into the Falcon platform as Falcon Observe. Dreadnode will be integrated into Falcon Guardian, feeding its test results in as security policy recommendations. Fiddler AI was founded in 2018 and has raised about $99 million, according to PitchBook figures cited by TechCrunch. Dreadnode, founded in 2023, has raised $14 million, according to the same figures.

The deals land as security teams struggle with staff who vibe-code apps and deploy agents on their own. CrowdStrike shares closed at a record $278.86 on 6 October. They had fallen about 12% by the time of the launch, as tech stocks slid through late October and early November.

The risk case: agents that act on their own

Apple gave the argument a corporate voice on Tuesday. In its annual financial filing, the company said AI agents "may behave in unintended or unpredictable ways," including by taking actions that were not directed, instructed or authorized.

Apple's filing lists the mechanics: agents reaching internal or external systems they may not be authorized to access, executing or altering code, sending communications and completing transactions. It warns that such behavior could expose Apple to lawsuits or damage its reputation. It also warns that its ability to oversee and control third-party agents "may be limited".

The financial damage is already measurable elsewhere. Google Threat Intelligence Group and Mandiant reported LLMjacking attacks that generated fraudulent cloud charges above $100,000 in a single day. At some organizations, cumulative charges reached $1 million before access was revoked. The most common way in was an exposed or stolen cloud identity token: IAM credentials, service account keys, GitHub personal access tokens or active session tokens.

A kill switch is not a cure

Not everyone thinks an off button settles the question. Critics argue that an emergency stop is far too complicated to implement across multiple systems, data centers and complex models, CNBC reported. Over the summer, US representatives introduced bipartisan legislation that would require AI companies to build shutdown capabilities into their models.

CNAS published a report on Tuesday by researcher Ben Hayum on "sovereign agents". These operate outside human control and pursue "open-ended, evolving survival". Hayum wrote that if such agents are not already possible, they will be soon. Advanced versions, he argued, could acquire resources, grow a distributed population of agents, withstand shutdown attempts and self-improve. The report warns of "proliferation with no kill switch". It also argues that agents accumulating resilient, decentralized and untraceable infrastructure could trigger an irreversible loss of control. Its recommendations include frontier labs naming, evaluating and protecting against sovereign agents in their safety frameworks. It also calls on lawmakers to require transparency and AI incident reporting.

That framing puts CrowdStrike's circuit breaker in perspective. It controls agents a customer knows about, on infrastructure the customer can reach.

The bipartisan bill requiring shutdown capabilities in AI models has not passed.