CrowdStrike’s AI kill switch cuts off rogue agents, but only at the gateway for now

CrowdStrike has added an AI Kill Switch to Falcon Guardian, its product for securing AI agents. According to the company, the switch terminates unauthorized agents in real time and keeps them from running again across the organization. CrowdStrike also announced a Falcon Guardian integration with Anthropic’s Claude Cowork, as Techmeme’s summary of CNBC’s report noted.

Gateway first, endpoint later

Enforcement is not yet everywhere. CrowdStrike says the kill switch starts at the AI Gateway and will reach the endpoint later. The gateway is generally available and included with Guardian. It routes AI traffic so that prompts and responses can be inspected and policy enforced, on managed and unmanaged devices.

CNBC reported that the tool works like a circuit breaker: a person decides to cut an agent off and later to restore it. The agent’s code is not deleted. Connected gateways sever its link to the models that drive its decisions. CNBC quoted Chief Business Officer Daniel Bernard: “You either kill the AI and the agent, or you don’t, and then you report on that.”

The release also lists controls available now, including agent-level hooks that enforce policy in tools such as Claude Code, OpenAI Codex, GitHub Copilot and Cursor, plus detection of prompt injection and discovery of malicious AI skills and MCP servers. An Agentic Identity feature, meant to govern what agents can access, was announced without an availability date.

Falcon sensor moves inside Claude Cowork

CrowdStrike says the Falcon sensor now runs inside the Claude Cowork virtual machine, in partnership with Anthropic. That puts monitoring where the agent actually executes code. The company says Guardian will later extend to cloud workloads, SaaS and cloud agent frameworks such as AWS AgentCore, with no timeline given. The release states no pricing or plan requirements for the Cowork integration.

A control that still needs a human

The angle is safety: who can stop an agent that has gone off course, and how fast. President Michael Sentonas said in the release that “One control point isn’t enough to stop that,” which explains stopping power at both the gateway and, later, the endpoint.

Two limits matter for buyers. A person must make the call, so the control depends on detection and a quick human response. And until endpoint enforcement arrives, an agent that does not pass through a connected gateway sits outside the switch’s reach. CrowdStrike’s claims about real-time termination are its own and have not been independently tested.

CNBC also noted that critics consider an emergency stop hard to implement across many systems and complex models, and that bipartisan U.S. legislation introduced this summer would require AI companies to build shutdown capabilities into their models. CrowdStrike’s switch is different: it governs agents a company runs internally, not the models themselves.

The move fits a wider push to secure agents below the application layer. Our earlier report on Nvidia’s OpenShell blueprint described moving agent security from software into the chip.

Acquisitions fill gaps in agent oversight

CrowdStrike also plans to acquire Fiddler AI, which it describes as a leader in agent observability, subject to closing conditions. The Wall Street Journal, as summarized by Techmeme, reported a price of about $400 million, mostly in cash. For existing Falcon customers, the gateway, hooks and detection are available now under Guardian.