CrowdStrike adds a kill switch for AI agents
CrowdStrike introduced an “AI Kill Switch” for Falcon Guardian to block rogue AI agents. The feature initially targets their access to models, with broader extension planned later.
CrowdStrike announced an “AI Kill Switch” for Falcon Guardian, a cybersecurity tool that allows authorized administrators to cut an AI agent’s access to AI models and providers. Presented on November 3, 2026, the feature targets agents capable of chaining tasks and acting with system rights on workstations, cloud environments, or SaaS applications. The issue is simple: if an agent misbehaves, it’s better to shut the door before it continues improvising.
Un contrôle humain pour des agents qui agissent trop vite
The innovation is based on centralized enforcement logic, meaning the application of rules that block the agent’s access to AI services. CrowdStrike specifies that an authorized administrator must trigger the shutdown and that every action is recorded in an audit log. The product addresses a concrete problem: agents can make unauthorized decisions without an external adversary necessarily being involved.
According to product documentation, the available response at launch goes through the Falcon AI Gateway, or through a third-party gateway already used in enterprises. CrowdStrike’s gateway is included at no extra cost with Guardian. It aggregates access to models, security policies, credential management, as well as usage and cost control within the Falcon platform.
Ce que le kill switch coupe vraiment, et ce qu’il laisse encore en chantier
The term “kill switch” can be misleading. At this stage, the announced control mainly cuts the agent’s connection to AI services; it does not yet neutralize the process itself across all environments. Direct termination functions on the workstation, blocking restarts, and revoking network access are on CrowdStrike’s roadmap, with a planned ramp-up later.
CrowdStrike also explains that Guardian combines several perspectives to avoid blind spots. Agent-level hooks see prompts but may miss some responses or actions; network tools alone do not always see local execution, nor open-weight models that do not transit the network. The platform thus mixes agent hooks, network inspection on the endpoint, and runtime visibility on the Falcon sensor.
Une sécurité pensée pour les usages agentiques en entreprise
Falcon Guardian has been generally available since September 2026, and the new control adds to a product already tasked with discovering AI agents in enterprises, applying policies, and detecting attacks targeting these agents. CrowdStrike Agent Graph links an agent’s instructions to its execution, as well as to identities, tools, systems, and resources involved. Security teams can thus reconstruct what happened without having to play digital archaeologists for three days.
The announced suite goes further: termination of cloud agents, revocation of execution permissions, restriction or shutdown of activity in SaaS environments, and browser session closure. Charlotte Agentic SOAR is to coordinate these actions and launch automated workflows based on AI-related detections or risky behavior. In the background, CrowdStrike advocates an idea increasingly adopted by the market: controlling an agent is not enough to monitor it; you must also be able to stop it.
Key points
- Presentation on November 3, 2026.
- Falcon Guardian generally available since September 2026.
- The kill switch blocks access to AI models.
- The Falcon AI Gateway is included at no extra cost.
- CrowdStrike Agent Graph traces instructions and execution.
- Charlotte Agentic SOAR is to orchestrate future blocks.
By the numbers
- November 3, 2026 — date of announcement, according to CrowdStrike.
- September 2026 — general availability of Falcon Guardian.
- 1 — authorized administrator required to trigger shutdown.