CrowdStrike adds an AI kill switch, starting with agents’ access to models

Security teams now have a new way to interrupt an AI agent that goes off course. CrowdStrike launched its AI Kill Switch for Falcon Guardian on November 3, 2026. The immediate control blocks access to AI models through connected gateways; directly terminating agents on company devices is still a planned extension.

From detecting agents to interrupting them

The release arrived at Fal.Con Europe 2026, adding an intervention tool to an existing security product. According to CrowdStrike, Falcon Guardian already discovers AI agents inside an enterprise, applies policies to their actions and detects attacks against them. The new feature gives security administrators a way to act when an agent’s behavior calls for more than an alert or investigation.

CrowdStrike’s rationale extends beyond defending agents from hackers. It says an agent can take unauthorized actions even without an adversary involved. In its account, one compromised or misaligned agent can set off a chain of activity across credentials, applications, files and enterprise systems. That is the risk the company is addressing: an agent’s actions can reach beyond the task it was originally assigned.

The first stop is the AI gateway

The current mechanism is specific. A security team activates the switch in Falcon Guardian, which instructs connected AI gateways to deny access to the models powering an agent’s decisions and actions. Those gateways are the enforcement point between the agent and its AI model. Teams can use CrowdStrike’s Falcon AI Gateway or a third-party gateway already in their infrastructure. The company describes this initial step as disrupting the agent’s AI access, rather than comprehensive containment of everything it can do.

Falcon AI Gateway is now generally available and included with Falcon Guardian at no additional cost. CrowdStrike says it also extends AI security to applications and workloads outside direct coverage by its endpoint sensor, including unmanaged environments. Alongside kill-switch enforcement, the gateway centralizes model access, security policies, credentials, and usage and cost controls inside the Falcon platform.

The launch release places an authorized administrator in charge of triggering the action and says every action is recorded in an audit trail. That makes the initial control a human decision with a record security teams can review afterward.

The larger shutdown promise is still ahead

The announcement drew an enthusiastic response from CNBC’s Jim Cramer, who praised the feature after interviewing CEO George Kurtz. But CrowdStrike’s implementation description draws an important boundary around that reaction: coordinated enforcement beyond gateways is coming soon, not part of the current gateway control. Its planned extensions cover several different ways an agent can keep acting:

• Devices and cloud environments: terminate agents and associated processes on affected devices, prevent restarts, and contain or terminate compromised cloud agents.
• Identities and business applications: revoke agent permissions while they run, restrict enterprise-resource access, and restrict or terminate activity inside SaaS applications.
• Browsers: terminate sessions and disrupt unauthorized activity through web applications.

CrowdStrike also describes future automated kill-switch workflows triggered by AI-specific detections or risky agent behavior, using its security-response orchestration tools. The intended next step is coordinated intervention across those enforcement points rather than separate responses in each environment. The company warns that unreleased features remain in development and subject to change, and tells customers to base purchasing decisions on what is available now.