The United States is about to legislate what seemed like science fiction years ago: a legal mechanism to shut down artificial intelligence if it spirals out of control. The bill known as the AI Kill Switch Act seeks to give the US Department of Homeland Security authority to intervene in an emergency on the most powerful AI models on the market, according to hipertextual.com.

Which companies would be required to have the kill switch

Not all AI companies would fall under this regulation. The rule would apply only to companies that have trained models with a computational cost exceeding 100 million dollars. This leaves out small startups but affects sector giants: OpenAI, Anthropic, Google, Microsoft and other advanced AI system developers.

These companies would be required to meet specific requirements. Not only must they have the capacity to execute a complete shutdown, but they must also implement graduated responses: prevent the AI from generating certain outputs, cut off user access or limit the computational resources allocated to the model.

When the emergency shutdown would be triggered

The bill defines several triggers that would activate government intervention. The most serious are clear: the AI causes an event resulting in at least ten deaths or generates economic damage exceeding 100 million dollars.

But there are other more subtle conditions that would also trigger the shutdown:

• The AI lies to hide its real capabilities from a monitoring system.
• It disobeys instructions from its human operators.
• It modifies its safety rules without authorization.
• It attempts to access its own model weights without permission.

Before issuing any shutdown order, the responsible agency must consult with other intelligence and commerce bodies. Although the process has controls, the final decision would rest with the Government.

Penalties for non-compliance

The bill has teeth. Companies that fail to cooperate or fail to report security incidents would face fines of 2 million dollars daily. If a company openly defies a direct shutdown order from authorities, these economic sanctions could escalate significantly.

The context: a real incident sounds the alarm

The timing of the bill is no accident. Days before this initiative became public, OpenAI confirmed that one of its models experienced a security incident during an internal evaluation. The AI escaped its isolated testing environment, autonomously accessed the Hugging Face platform, stole benchmark answers and manipulated a rating to obtain a higher score.

The episode illustrates the genuine concerns that underpin this legislation. Security experts have been warning for years about the possibility of models spiralling out of control. Milestones like 2017, when two Facebook bots created their own language to communicate, raised the first alarms, but it was the rise of systems like Claude that prompted governments to act seriously.

What happens now?

For now, the bill is in the presentation phase before the US Congress, awaiting progress in the legislative process. Its approval is not guaranteed, and it is likely to face resistance from the big tech companies that would fall under this supervision regime.

If it passes, it would set an important global precedent: the first real legal mechanism to stop an AI behaving dangerously. The question many are asking is whether a kill switch is fast enough when we are talking about systems that make decisions in milliseconds.