CrowdStrike has launched an AI Kill Switch that lets security teams shut down unauthorized AI agents, but the first version works only through a network gateway. Blocking agents directly on computers, known as endpoints, is planned for a later release with no set date.
The cybersecurity company announced the feature on Tuesday, November 3, at its Fal.Con Europe 2026 conference in Barcelona. It also signed an agreement to buy Fiddler AI, a startup that tracks and explains how AI agents behave.
The new tools expand Falcon Guardian, CrowdStrike’s platform for securing AI agents. Guardian has been generally available since September 2026.
What Works Today and What Comes Later
The kill switch currently operates at the AI Gateway, which can cut off an agent’s access to the models that power its decisions. CrowdStrike said the tool terminates unauthorized agents in real time and blocks them from running again across an organization.
Only an authorized administrator can trigger the kill switch. Each use is recorded in an audit trail.
CrowdStrike plans to extend the feature so it can end agent processes directly on endpoints. The company also described future controls that would coordinate restrictions across cloud environments and other enterprise systems.
The AI Gateway itself is now generally available and included with Guardian. It routes AI traffic through one central point to inspect prompts and responses, track usage, and enforce policy on managed and unmanaged devices.
Endpoint network inspection, another new feature, checks AI traffic as it enters and leaves a device. It is generally available for Windows, while the Mac version is in early access.
Guardian also uses agent-level hooks to apply policy inside popular AI tools. These include Claude Code, OpenAI Codex, GitHub Copilot, Microsoft 365 Copilot, Cursor, and OpenCode.
The platform uses CrowdStrike’s SafeMind models to detect attacks on AI systems, such as prompt injection and malicious instructions. A separate supply chain feature looks for harmful AI skills and MCP servers.
Other features are still in development. Identity-aware controls that decide what agents can access are planned, along with expanded coverage for cloud workloads, SaaS apps, Claude Cowork, and AWS AgentCore.
Michael Sentonas, president of CrowdStrike, said “every week we see agents chain steps together, hit an obstacle, and take actions they were never authorized to take.” He said stopping that requires protection at the agent, the network, and the endpoint together.
CrowdStrike is also working with Anthropic to run its Falcon sensor inside the Claude Cowork virtual machine. The setup gives security teams visibility into the Cowork environment.
The launch follows a flaw that Microsoft researchers found in Anthropic’s Claude Code GitHub Action. The issue could have let attackers trick the agent into exposing credentials, and Anthropic has since patched it.
Fiddler Deal Expected to Close by January
Fiddler’s software tests agents before they are deployed, traces their decisions and tool calls, and monitors changes in behavior over time. These records can help security teams investigate unexpected actions or provide evidence to auditors.
The Wall Street Journal reported that the deal values Fiddler at about $400 million. CrowdStrike has not disclosed the price but said it will pay mostly in cash, with some stock subject to vesting conditions.
CrowdStrike expects the acquisition to close in the fourth quarter of its fiscal 2027, which ends January 31, 2027. The deal remains subject to customary closing conditions.
