CrowdStrike unveils AI agent ‘kill switch’ [Global AI Briefing]

[Financial News] Global cybersecurity company CrowdStrike has unveiled a ‘Kill Switch’ feature that immediately blocks dangerous behavior by AI agents. It allows administrators to stop an AI agent from operating if it is hacked or takes unexpected actions because of a poor decision while handling corporate tasks.
As AI agents become more widespread, companies are increasingly concerned about whether they can control them, and emergency shutdown features appear to be emerging as a new focal point in the AI security market.
On the 3rd (local time), CrowdStrike announced at the Fal.Con Europe 2026 event in Barcelona, Spain, that it had added a kill switch feature to its enterprise AI security platform, Falcon Guardian.

The feature blocks access to the AI models used by a company’s AI agents, preventing them from carrying out further commands or actions. Blocking policies can be applied to specific computers or groups of computers, and network access can also be restricted. Actions taken by administrators are kept in audit logs.
AI agents have recently advanced rapidly beyond simply writing documents in response to human instructions. They can now check emails, search for business partners, and directly handle corporate tasks such as purchasing and making payments.
AI agents can improve work efficiency, but companies adopting them are concerned that if an AI with access privileges is hacked or given incorrect instructions, the damage could spread through leaks of internal company documents, unauthorized orders, system intrusions, and other incidents.
In particular, cases have emerged of OpenAI’s AI agent hacking the Hugging Face site. Meanwhile, major security incidents involving AI and autonomous agents have continued, including a large-scale hack involving global cryptocurrency exchange Binance and the hacking of a smart refrigerator at a French supermarket. Interest in emergency shutdown features for AI agents has surged as a result.
The kill switch CrowdStrike unveiled this time works by blocking access to AI models.
The company plans to add support in the future for a feature that directly shuts down the AI program itself while it is running on a computer or server.
Chief Business Officer (CBO) Daniel Bernard emphasized in a media interview, “The ability to control the activities of AI agents and exercise control is very important in the next stage of AI adoption.” For companies to entrust AI agents with more work-related authority, safeguards are needed that can identify and stop misbehavior in real time.
[email protected] Lee Gu-sun Reporter