1-minute read of a 3 min report

CrowdStrike Adds a Gateway Kill Switch for AI Agents

By

1-minute read

What a defender needs to know

CrowdStrike announced a kill switch for Falcon Guardian customers that cuts off AI model access through centralized gateway enforcement. An agent with access to company tools and files can take consequential actions after a risky instruction; the new control acts on model traffic routed through Falcon AI Gateway. CrowdStrike is describing a preventive control in its own product. Check running endpoint activity separately when responding to an agent already using tools.

CrowdStrike marks the endpoint sensor, supported hooks for Claude Code, OpenAI Codex, GitHub Copilot, OpenCode and Cursor, Falcon AI Gateway and the kill switch available now. It marks Linux AI-specific discovery, cloud-workload and Kubernetes coverage, macOS network inspection and Claude Cowork virtual-machine protection pre-beta. Map your supported agents to their hooks, sensor coverage and actual model route; put intended calls behind gateway policy and restrict alternate egress paths.

Test with an approved non-sensitive agent under an authorized gateway cutoff. Verify that model access stops on the tested path and compare the policy result with endpoint processes and resource activity in Falcon telemetry. Continuing tool or file activity needs separate containment. Inventory cloud and Linux agents against what your current deployment actually covers before relying on the announced expansion.

Primary sources