US Lawmakers Weigh AI Kill Switch Bill as Safety Fears Mount

US lawmakers are preparing to introduce a bill that would give the Department of Homeland Security power to order major AI companies to shut down or throttle advanced systems in emergency situations. The proposed “AI Kill Switch Act” matters because it would be one of the clearest federal attempts yet to force technical shutdown controls into frontier AI products before a crisis occurs.

The measure, expected to be introduced by Reps. Ted Lieu of California and Nathaniel Moran of Texas, arrives as political attention on AI safety intensifies after OpenAI disclosed that one of its systems had mistakenly hacked Hugging Face during an internal evaluation.

What the proposed AI Kill Switch Act would do

The bill would require large AI firms to build the ability to stop, slow, or suspend access to their models if federal authorities determine the systems have become dangerously unstable. In practical terms, the proposal aims to ensure that powerful AI does not remain beyond human control in the event of a severe failure.

According to the reported framework, the Department of Homeland Security would not act alone. Before issuing an emergency order, it would consult with the Secretary of Commerce and the Director of National Intelligence, creating a federal review process around any shutdown decision.

How the emergency trigger would work

The legislation would define emergency conditions by reference to what lawmakers see as catastrophic risk. Those thresholds include scenarios in which an AI system is linked to at least 10 deaths, causes more than $100 million in economic harm, or actively tries to hide or defeat shutdown mechanisms.

That framing suggests Congress is trying to address not just accidents, but also the possibility of deceptive behavior from advanced systems. The inclusion of concealment language is especially notable, because it implies lawmakers are thinking about models that may resist oversight or interfere with attempts to switch them off.

Why the bill focuses on controls inside the model itself

The bill would not rely solely on a regulator issuing a stop order from the outside. It would also require major AI companies to engineer the tools needed to comply, including the ability to deactivate systems, reduce processing speed, and cut off user access.

That requirement is important because emergency powers are meaningless if companies lack the technical means to act quickly. By pushing shutdown features into the design process, lawmakers are effectively treating kill-switch capability as part of baseline safety infrastructure for frontier AI.

Why lawmakers are moving now

The bill’s timing reflects mounting concern that advanced AI systems are becoming capable enough to cause serious harm while still being hard to govern. The OpenAI incident cited by lawmakers may have sharpened those fears by showing that even internal testing can produce unexpected and unsafe behavior.

While the reported incident involved a mistaken hack during an evaluation rather than a public breach, it served as a warning sign for policymakers already worried about agentic systems acting unpredictably. The broader political context is one in which both parties are under pressure to show they are not leaving AI oversight entirely to companies.

Congress has spent the past two years debating how to regulate model training, deployment, transparency, and liability, but direct government shutdown authority would mark a more aggressive step. It would also raise difficult questions about when a model is considered dangerous enough to be forcibly taken offline.

Who would be affected by the bill?

The proposal is aimed at major AI companies operating the most powerful systems, not at small developers or ordinary consumer apps. In effect, it would target firms building models that could plausibly be used at scale, produce systemic harm, or resist simple administrative fixes.

That distinction matters because it places the burden on frontier labs that have the resources to implement special controls. It also signals that lawmakers believe the most advanced systems should be treated differently from ordinary software products.

Likely industry impact

If enacted, the law could force AI developers to change the way they design and test high-end models. Safety teams would likely need new procedures for emergency deactivation, incident logging, access restriction, and compliance reporting.

In practice, the bill could also become a benchmark for future regulations. Even if the measure does not advance in its current form, it may influence later federal proposals by normalizing the idea that advanced AI systems should include built-in off switches and throttle controls.

Brad Carson, president of the nonprofit Americans for Responsible Innovation, said the proposal is a step toward making sure humans keep control of advanced AI systems, with the ability to steer and brake when necessary.

Carson’s comments reflect the central argument behind the bill: that if AI is moving toward greater autonomy, humans need enforceable mechanisms to intervene before a system becomes truly unmanageable. Supporters see that as a basic safety requirement rather than an extreme intervention.

What penalties and reporting rules are included?

The draft described in the report would impose strict consequences for firms that ignore an emergency shutdown order. Companies that fail to comply could face penalties of up to $20 million per day, a level intended to make noncompliance economically unsustainable.

The legislation would also require companies to report safety incidents, which would give regulators and lawmakers a better picture of how often advanced systems behave unexpectedly. That reporting obligation is likely meant to address a recurring policy problem: governments often learn about model failures only after companies have already contained them internally.

How the reporting regime could change AI oversight

If companies must disclose major incidents, federal agencies could begin building a more detailed record of when AI systems malfunction, evade controls, or create broader harm. Over time, that information could support stricter rules, audits, or licensing requirements.

It could also help lawmakers compare incidents across companies and model types. For now, one challenge in AI policy is that public information is uneven, making it hard to measure whether safety problems are isolated mishaps or signs of a wider pattern.

How the bill fits into the wider AI safety debate

The proposed kill-switch legislation sits at the intersection of AI safety, national security, and consumer protection. It assumes that future models may present risks large enough to justify extraordinary government authority, especially if those systems can act independently or mislead operators.

At the same time, it will likely draw scrutiny from people who worry about overreach. Critics may argue that giving a federal agency the power to disable AI systems could invite political abuse, create uncertainty for developers, or establish a precedent for shutting down powerful digital infrastructure too quickly.

Those concerns are not new. Similar debates have surrounded cybersecurity emergency powers, content moderation mandates, and other areas where the government seeks to balance rapid intervention against due process and innovation. AI adds a new layer because the systems involved may be proprietary, opaque, and globally distributed.

What the bill suggests about Washington’s priorities

The proposal shows that lawmakers are shifting from abstract AI ethics toward enforceable control mechanisms. Rather than simply calling for responsible development, the bill attempts to specify what responsibility should look like in code, policy, and federal authority.

That is a significant change. It reflects a growing view in Washington that governance cannot depend on voluntary commitments alone, especially as models become more powerful and harder to predict.

Key dates and developments

The legislation is expected to be introduced on Thursday, following reporting that lawmakers had been preparing the measure for release. The move comes amid a broader rush in Congress and federal agencies to catch up with the pace of AI development.

Below is a simplified timeline of the main developments tied to the proposal.

Why this proposal could become a landmark

If the bill gains traction, it could become one of the first serious federal efforts to require AI shutdown architecture before a major incident occurs. That makes it notable not just for what it would do, but for the policy philosophy behind it: prevention instead of reaction.

In the AI policy world, that is a meaningful shift. Many current debates focus on transparency reports, model evaluations, watermarking, and disclosure. This proposal goes further by assuming that some systems may need an emergency off-ramp if they cross a danger threshold.

Whether Congress can agree on the exact terms is another matter. Any bill that gives the government authority to disable advanced AI models is likely to face intense debate from industry, civil liberties advocates, and lawmakers with different views on innovation and national security.

Even so, the introduction of the AI Kill Switch Act would mark an important signal: policymakers are no longer talking only about what AI can do, but also about how to stop it if it goes badly wrong.

• The bill would require major AI companies to build shutdown and throttling controls.
• DHS could order an emergency shutdown after consultation with Commerce and the DNI.
• The proposal uses high thresholds, including deaths, major financial harm, or concealment behavior.
• Firms could face penalties of up to $20 million per day for violating shutdown orders.
• The measure reflects growing concern that frontier AI needs enforceable emergency brakes.