CrowdStrike adds AI kill switch to Falcon Guardian to stop rogue agents

A kill switch for agents that misbehave

According to CrowdStrike, the new AI kill switch is available now to Falcon Guardian customers. It is designed to let security teams rapidly cut off AI model access at AI gateways once an agent becomes dangerous.

The company frames the feature as a response to a shift in enterprise computing. In its words, what began with employees using AI assistants is evolving into autonomous agents that access enterprise resources, invoke tools, execute code and take action at machine speed. Organizations, CrowdStrike says, must protect those agents from attacks while making sure they operate safely and as intended.

CrowdStrike also announced the general availability of Falcon Complete for Guardian, which extends 24/7 managed detection and response to AI security. The announcements advance what the company calls AI detection and response (AIDR) across three dimensions: architecture, coverage and controls.

Three layers of visibility

CrowdStrike argues that traditional approaches each show only part of the picture. Endpoint runtime security monitors execution but lacks the AI-specific context to explain what triggered it. Network inspection can monitor AI communications but not the actions agents take on systems. Agent-level integrations can inspect AI interactions but cannot reveal downstream execution.

Falcon Guardian is built to combine three layers so that security teams can connect what agents are instructed to do with what they actually execute:

• OS-level runtime security (available now): the Falcon sensor provides visibility into the processes agents launch, the commands they run and the resources they access. CrowdStrike says this visibility is not achievable directly through network inspection or per-agent hook integrations.
• Endpoint network inspection (pre-beta): CrowdStrike announced plans to expand this capability to macOS, where it will inspect traffic entering and leaving the endpoint to help teams identify risky interactions.
• Agent-level hooks (available now): hooks for Claude Code, OpenAI Codex, GitHub Copilot, OpenCode and Cursor let organizations inspect prompts and block malicious activity at the source.

What it means for security teams

The announcement lays out what CrowdStrike sees as the tasks security teams now face with agents: discovering where agents operate, understanding what they were told to do, tracing their actions and intervening when behavior turns dangerous. The kill switch addresses the last of these, while the runtime and hook layers address visibility and tracing.

The post's summary also says the release strengthens AI security architecture, expands cloud coverage and enables agent insider threat detection, though the portion of the source available here does not detail those cloud and insider threat capabilities. Several features are at different stages: the runtime sensor and agent hooks are available now, while macOS network inspection is still pre-beta.

The same day, CrowdStrike published related AI security news on its blog. Its listed titles include plans to acquire Dreadnode to bring automated AI security testing to the Falcon platform, plans to acquire Fiddler to deliver Falcon Observe for AI agent observability, and a post on securing AI agents inside Claude Cowork. Together with the kill switch and managed detection and response offering, they show the company building out controls for organizations that are letting autonomous agents act inside their systems.