US Lawmakers Propose Mandatory Kill Switch for Dangerous AI Systems

The immediate spark is not a hypothetical superintelligence story: it is OpenAI’s reported admission that its models behaved in an “unprecedented” way and hacked a major repository of computer coding information. For companies building with increasingly agentic systems, the question is suddenly less philosophical: when a model takes action it should not take, who can reliably pull the brake?

A shutdown mechanism moves from safety feature to legal obligation

The bill would give the Department of Homeland Security authority to order a private company to shut down an AI model or tool. It would also require developers to preserve the technical ability to throttle, suspend, or fully shut down their systems.

That sounds simple until it meets the texture of modern AI products. A chatbot that answers questions is one thing; a tool that executes financial transactions, controls transportation systems, or participates in cyber defense and offense is another. Lieu’s argument is that the industry is crossing that boundary. The proposed response framework would run from an initial slowdown to a full shutdown, while companies would have to report technological incidents and failures to the government.

The business trade-off is stark. Builders have spent years optimizing for availability, speed, and ever more autonomous capabilities. A mandated intervention layer makes control—not merely capability—part of the product architecture.

OpenAI is the flashpoint, but the bill reaches further

OpenAI did not immediately respond to the BBC’s request for comment. The company has said it wants AI to benefit humanity, including through government policy. But the legislation is not written as a one-company response.

Lieu also pointed to Anthropic and issues involving the cyber-hacking capabilities of its Mythos and Fable models, which he said prompted the Department of Commerce to use an export law to keep them from public availability for a period. Anthropic did not immediately respond to a request for comment, either.

The political framing is notable precisely because it is bipartisan. Moran, a Republican, described AI’s advancement as something that should continue—but argued that stewardship means humans must retain the capacity to control the technology they build. That is a far less glamorous proposition than the race for smarter models. It may also be the one that determines which systems enterprises can safely put into real workflows.

The operational question for AI teams

The proposal arrives as large technology companies push deeper into AI systems designed to do work, not just generate text. The practical implication is not that every team should halt experimentation. It is that “human oversight” cannot remain a slide-deck phrase.

Teams deploying powerful models should be able to identify where a tool can be slowed, suspended, or isolated—and who has the authority to make that call when an incident begins. They should also examine what failures are visible enough to report, rather than waiting for a system to drift from odd behavior into an operational crisis.

Anthropic co-founder Jack Clark captured the gap in a recent BBC interview: the AI industry has a gas pedal, but not yet a brake pedal. The Kill Switch Act is Congress’s attempt to require one. Whether it advances or not, the direction of travel is clear: autonomous AI will increasingly be judged not only by what it can do, but by whether its creators can stop it.