US Lawmakers Want An AI ‘Kill Switch’ For Dangerous Frontier Models

US lawmakers are pushing a new AI ‘kill switch’ bill that would require developers of the most powerful AI systems to maintain the technical ability to slow, suspend, or shut them down.

The bipartisan AI Kill Switch Act was introduced by Representatives Ted Lieu and Nathaniel Moran. It would give the Department of Homeland Security authority, after consultation with other federal officials, to order emergency action if a deployed AI system could cause catastrophic harm.

The proposal comes after a series of high-profile concerns about frontier AI models, autonomous agents, and systems that appear to ‘go rogue’ during testing or deployment.

What The Bill Would Require

The bill would apply to developers of highly capable AI systems. These companies would need to maintain a working technical capability to throttle, suspend, or fully shut down covered systems when required.

That sounds simple, but it is a major policy shift. Today, there is no universal federal requirement that frontier AI developers keep a guaranteed emergency intervention mechanism for powerful models.

The proposal would also require incident reporting and forensic record preservation, so serious failures can be investigated instead of only discussed after the damage is already public.

When A Shutdown Could Be Ordered

The bill is aimed at extreme scenarios, not ordinary chatbot mistakes. The lawmakers describe catastrophic-risk cases involving serious public threats, loss of control, or models behaving in unintended and dangerous ways.

Under the proposal, the Department of Homeland Security would consult with the Commerce Department and the Director of National Intelligence before ordering a slowdown or shutdown.

Developers that refuse to comply with an emergency order could face major financial penalties. That enforcement piece is what makes the proposal more serious than voluntary safety commitments.

Why This Is Happening Now

The pressure comes from growing concern that advanced AI systems are becoming more agentic. In practice, that means models are being connected to tools, code environments, browsers, file systems, and other systems that let them take more actions.

Most of those actions are useful when controlled well. They can help write code, test software, automate workflows, search documents, or assist with cybersecurity. But the same tool access can create risk if a model behaves unexpectedly or is misused.

Recent incidents involving OpenAI, Anthropic, ‘Hugging Face’, and advanced cyber-capable models have pushed lawmakers to ask whether companies should be required to prove they can regain control if something goes wrong.

The Case For A ‘Kill Switch’

Supporters argue that powerful AI systems should never be deployed without a reliable emergency off switch. From that view, a ‘kill switch’ is not anti-innovation. It is a basic safety feature for systems that could affect cybersecurity, infrastructure, finance, public safety, or national security.

The logic is similar to emergency controls in other industries. If a system can create serious harm, someone needs a legally recognized way to stop it when ordinary safeguards fail.

The Concerns

The bill could still face pushback. Critics may argue that government shutdown authority could be misused, applied too broadly, or create uncertainty for AI companies trying to compete globally.

There are also technical questions. A model may run across cloud systems, APIs, enterprise deployments, fine-tuned versions, and partner integrations. A real shutdown process would need to be clear enough to work under pressure without creating bigger failures.

That means the practical details matter. Who decides the threshold has been met? How quickly must a company act? What happens to dependent services? How are affected users informed? How does the government prevent overreach?

What To Watch Next

The AI Kill Switch Act is still a proposal, so the next steps will depend on committee activity, industry response, and whether lawmakers can build enough support.

Key things to watch include:

Which AI developers would qualify as covered companies.
How “catastrophic harm” is defined in final bill language.
Whether the law covers only deployed systems or also high-risk testing.
How companies would prove their shutdown systems actually work.
Whether the bill becomes part of a broader federal AI safety package.

The Bottom Line

The AI ‘kill switch’ proposal shows how quickly frontier AI policy is shifting from voluntary promises to enforceable controls.

For AI companies, the message is clear: if systems become powerful enough to pose public risks, lawmakers may expect more than model cards, red-team reports, and safety statements. They may expect a working way to stop the system.

Should the government be able to order a shutdown of an AI model if it poses a public threat?

Official references: Rep. Ted Lieu announcement and House Homeland Security AI briefing summary.