House lawmakers are set to introduce legislation that would give the Department of Homeland Security (DHS) the power to order AI companies to shut down or throttle their systems, with penalties of up to $20 million per day for refusing to comply. Representatives Ted Lieu (D-CA) and Nathaniel Moran (R-TX) plan to file the AI Kill Switch Act on Thursday, marking the first serious federal attempt to codify what emergency AI safety controls should look like at the infrastructure level.

What Would Trigger a Government AI Shutdown?

The bill establishes specific, concrete conditions that would allow DHS to exercise shutdown authority. The trigger thresholds are designed to set a high bar, preventing the government from ordering shutdowns over policy disagreements or minor incidents. Under the proposal, DHS could act in consultation with the Secretary of Commerce and the Director of National Intelligence, creating a three-agency structure that keeps shutdown authority out of the hands of a single official.

• Death Threshold: At least 10 deaths directly caused by an AI system
• Economic Damage Threshold: More than $100 million in economic damages resulting from an AI incident
• Concealment Trigger: A model attempting to hide or disable its own shutdown controls

The timing of the bill is not coincidental. It follows OpenAI's disclosure that its systems mistakenly hacked Hugging Face, a popular platform for sharing AI models, during an internal evaluation. Security researchers determined that a sandbox misconfiguration, rather than a rogue model, enabled the breach, but the incident provided lawmakers with a concrete example of how AI systems can behave unpredictably.

How Would Companies Comply With Shutdown Orders?

The legislation would impose an engineering mandate on frontier AI developers, requiring them to build specific technical capabilities into their systems. Companies would need to design their models with the ability to respond to government commands in real time, creating what amounts to a federal off-switch built into the architecture of advanced AI systems.

• Shutdown Capability: The ability to completely halt a model's operations on command
• Throttling Controls: The ability to slow down or reduce a model's performance and capabilities
• User Access Suspension: The ability to suspend user access to a system without taking it offline entirely
• Safety Reporting: A requirement to report safety incidents to the government, creating a formal disclosure regime that does not currently exist at the federal level

The $20 million per day penalty is the sharpest enforcement mechanism in the bill. For frontier AI developers generating hundreds of millions of dollars in quarterly revenue, sustained non-compliance could accumulate to nine figures within a week. The figure is calibrated to make ignoring a DHS order economically irrational rather than serving as a minor nuisance fine.

What Is the Industry's Likely Response?

Support from outside groups is already forming. Brad Carson, president of Americans for Responsible Innovation, a nonprofit focused on AI governance, framed the bill as a necessary guardrail rather than a brake on AI development.

"This is an important step toward ensuring that humans have both hands firmly on the wheel and a foot ready at the brake as advanced AI systems are deployed," stated Brad Carson.

Brad Carson, President of Americans for Responsible Innovation

Industry pushback is expected. AI developers have generally argued that shutdown authority in the hands of a federal agency introduces political risk, competitive risk, and the possibility that a shutdown order itself becomes harmful, cutting off critical systems like hospital diagnostic tools, bank fraud detection layers, or classified defense contractor deployments. The bill's high threshold language of 10 deaths and $100 million in damages is designed to address these concerns, but the specific procedure for lifting a shutdown order is where the real fight will occur if the bill advances.

The proposal arrives in a Congress that has debated AI safety legislation repeatedly without passing substantial measures. The bipartisan pairing of Lieu and Moran gives the AI Kill Switch Act better odds than a single-party bill, but the path from introduction to floor vote for AI legislation has historically been slow. The more likely near-term effect is to shape the terms of the broader debate, specifically establishing the idea that the federal government should have a defined off-switch rather than negotiating one incident at a time.

For the AI industry, the bill matters even if it never becomes law. A federal shutdown authority with a $20 million per day penalty becomes the baseline that any voluntary safety commitment, any state-level bill, and any international framework gets measured against. Companies that already ship shutdown and throttling controls have a compliance story to tell; companies that do not will find themselves engineering against a specification written in Washington rather than one of their own choosing. The AI Kill Switch Act represents the first serious attempt to codify what human oversight actually means at the infrastructure layer, and the enforcement number attached to it will pull the entire policy conversation toward a federal standard.