CrowdStrike launches an AI kill switch for rogue agents as it chases $6.6B in recurring revenue

CrowdStrike on Tuesday launched what it calls the AI Kill Switch, a control inside its Falcon Guardian product that lets a security team stop an AI agent that has gone off course, keep it from launching again and cut its access to the corporate network. The company says it is available today. CNBC reported the launch the same afternoon.

The headline is bigger than the first version of the product. Read the release closely and the "kill" part is staged.

What it actually does

According to CrowdStrike's release, the switch has four parts:

• Terminate the running agent. The Falcon sensor stops the agent from executing, "first at the Gateway, and soon on the endpoint."
• Block it from running again. A Falcon Guardian policy keeps the agent from relaunching on the targeted machine or group of machines.
• Cut its network access. The agent loses its enterprise network access.
• Keep a human in charge. An authorized administrator triggers the action, and every action goes into an audit trail.

CNBC adds two details from CrowdStrike. The tool does not delete the agent's code. It uses connected AI gateways to cut off the agent's access to the AI models that drive its decisions. And it needs a person: the system can detect internal agent activity, but a human decides to cut an agent off and to restore it. Chief Business Officer Daniel Bernard described it to CNBC as a circuit breaker and said that when you kill a process at the root, "it's very binary."

Falcon Guardian itself launched in September to discover AI agents running in a company, enforce policy on what they can do and detect attacks against them, per the release.

Who is on record using it

One customer is quoted: Rick Orloff, chief information security officer at Everpure, who said the switch "runs on the Falcon sensor we already have." That is a statement about convenience for an existing CrowdStrike customer, not a report of an agent being stopped in production. The release names no other customers and gives no pricing, and we found no published incident in which the switch contained a real rogue agent.

Why the timing

CNBC ties the launch to a run of incidents, including a Hugging Face breach in July that it says was carried out by rogue OpenAI agents, and to bipartisan legislation introduced this summer that would require AI companies to build shutdown capability into their models. CNBC also notes critics who argue an emergency stop is too complicated to implement across many systems, data centers and models. Note the difference: that bill is about the AI labs shutting down their own models. CrowdStrike's tool is a customer-side control over agents running inside one company's own systems.

The numbers behind the pitch

CrowdStrike is selling this into a business that is already growing. In its second-quarter fiscal 2027 results, filed with the SEC on Aug. 26 for the quarter ended July 31, the company reported:

Measure	Reported or guided
Total revenue, Q2	$1.47 billion, up 26%
Annual recurring revenue (ARR), July 31	$5.84 billion, up 25%
Net new ARR, Q2	$332.8 million (a record)
Free cash flow, Q2	$377.4 million
GAAP net income, Q2	$5.3 million (non-GAAP $322.9 million)
FY27 guidance: revenue	$5.991 billion to $6.011 billion
FY27 guidance: ending ARR	$6.603 billion to $6.612 billion
Q3 guidance: revenue	$1.523 billion to $1.529 billion

The midpoints of the ARR guidance imply roughly $0.35 billion of net new ARR in the October quarter and about $0.42 billion in the January quarter. That is our arithmetic on the company's guidance, not a company figure, and guidance is a forecast, not a result. CEO George Kurtz said in the release that securing AI "is the largest market opportunity in our history."

Note the GAAP line. The company's profit under standard accounting was $5.3 million on the quarter; the large profit figures are non-GAAP, which exclude items such as stock-based compensation. Both are in the filing. We also reported in September that the DOJ had closed its probe, while the SEC's review of ARR reporting was not publicly resolved, which matters because ARR is the metric this launch is meant to grow.

The deals announced alongside it

The same day CrowdStrike said it signed definitive agreements to buy Fiddler AI, for agent observability, and Dreadnode, for automated AI security testing. The company did not disclose prices. Stocktwits, citing the Wall Street Journal, put the total at $495 million (about $400 million for Fiddler and $95 million for Dreadnode); we have not seen the Journal report ourselves, so treat that as unconfirmed. Per that account, both deals are funded mainly with cash plus some stock and are expected to close in CrowdStrike's fiscal fourth quarter of 2027. CNBC also reported an integration with Anthropic's Claude Cowork.

The competition

CrowdStrike is not first with the phrase. JetStream Security announced its own AI Kill Switch on July 27, which it says shuts down one compromised agent in seconds without disturbing the rest of a company's AI. Other security vendors have also advertised agent shutdown controls, but we did not verify those products. The relationship is not purely rivalry: CrowdStrike's August results list JetStream among the AI gateway partners its AI Detection and Response product works with, and the kill switch relies on connected gateways.

What is unproven

• Endpoint termination. The most direct form of the kill, stopping the agent on the machine, is described as coming later. No date is given.
• Real-world performance. We found no independent test, no stated stop times and no named case of the switch halting a rogue agent in production.
• Customer adoption. One quoted customer. CrowdStrike gave no count of companies using Falcon Guardian.
• Revenue impact. The company has not said what the kill switch adds to ARR or whether it is a separate charge. Its acquisitions are not due to close until the fiscal fourth quarter, and the related new capabilities are also scheduled for then, per Stocktwits.
• Human in the loop cuts both ways. Requiring an administrator to pull the switch is a safety feature, and also a delay if an agent acts faster than a person can respond.

What it means for a business running agents

For a company already on the Falcon sensor, this is another control to evaluate. For any business letting AI agents touch email, payments or customer data, the practical question is the one the release implicitly raises: can you list your agents, and can someone with authority stop one tonight? That is true whichever vendor you use.

Sources: CrowdStrike press release; CNBC; CrowdStrike Q2 FY27 results, SEC EDGAR; JetStream Security. Chart: The Company Chronicle, from CrowdStrike's results release. This is market information, not investment advice.