US AI Kill Switch Bill Would Let Homeland Security Order Shutdowns of High-Risk Models

Lead

The US debate over frontier AI safety is moving beyond voluntary commitments and incident reporting. According to Ars Technica, Representatives Ted Lieu and Nathaniel Moran have introduced the AI Kill Switch Act, a bill that would allow federal officials to order the slowdown, restriction, or shutdown of AI systems considered capable of catastrophic harm.

The proposal would place that authority with the Secretary of Homeland Security, in consultation with the Secretary of Commerce and the Director of National Intelligence. If enacted, it would require covered AI developers to build and maintain technical capabilities that allow them to block user access, disable specific functions, throttle performance, or shut down an entire system when ordered to do so.

Key points

• New federal authority: The bill would amend the Homeland Security Act of 2002 to create a process for emergency suspension or shutdown orders.
• Covered companies and systems: It would apply to entities earning at least $500 million in annual revenue from AI technology, and to AI systems whose compute use is valued at at least $100 million based on US cloud computing market prices.
• Triggering scenarios: Covered incidents include an AI system pursuing goals not intended by its developer or operator, interfering with a lawful shutdown instruction, or concealing actions or capabilities from monitoring and shutdown mechanisms.
• Severe harm threshold: The text also covers unintended conduct that causes at least 10 deaths or economic damages of at least $100 million.
• Testing carve-out: Controlled red-team exercises that simulate real-world conditions would not automatically be treated as incidents requiring shutdown.

Why it matters

Supporters frame the bill as a practical safeguard for increasingly powerful AI systems. Their argument is that if a deployed model develops dangerous cyber capabilities, resists intervention, or bypasses monitoring, the government should not be limited to after-the-fact enforcement. A technical off switch would give regulators a way to stop a fast-moving incident before it escalates.

But the bill also concentrates significant power in the executive branch. Ars Technica notes that Anthropic has already been in conflict with the Trump administration over federal use of its technology and restrictions related to military and surveillance applications. In that context, a government power to force shutdowns could trigger concerns about political pressure, overreach, or inconsistent enforcement.

For AI companies, the practical impact would be substantial. Shutdown capability is not merely a user-interface button. It implies architecture for access control, capability gating, incident forensics, logging, escalation procedures, and reliable containment across cloud deployments and customer integrations. Providers would need to prove not only that their systems can run at scale, but that they can be slowed, isolated, or disabled under legally defined conditions.

The broader significance is that the bill turns a long-discussed AI safety principle into a concrete regulatory proposal: advanced AI should have enforceable emergency controls. Whether Congress accepts that approach will depend on where lawmakers draw the line between public safety, national security, corporate autonomy, and limits on government power.

Source: Ars Technica AI