Jarred by a recent rogue AI incident, a bipartisan proposal in the US House of Representatives seeks to give the federal government the power to order companies to slow, suspend, or shut down powerful artificial intelligence systems if they escape human control or pose a risk of catastrophic harm.
The AI Kill Switch Act, introduced on Thursday by Democratic Congressman Ted Lieu and Republican Congressman Nathaniel Moran, comes amid growing concern that rapidly advancing AI agents can do more than generate text and images. They can also independently execute computer code, conduct financial transactions, and interact with critical systems.
But the proposed "kill switch" would not be a single button in a government building capable of instantly turning off any AI system. Instead, the legislation would require major developers to build and preserve the technical ability to disable their own models and obey emergency orders issued by the US Department of Homeland Security.
The bill would apply only to some of the most powerful and commercially significant AI systems.
A "covered technology" would generally be an AI system developed using computing power that would cost more than $100 million at prevailing US cloud-computing prices.
For a company to qualify as a "covered entity," it would also need to operate the system, make it available to outside users through an application programming interface or hosted service, and generate at least $500 million in annual gross revenue from the technology. Personal, academic, and non-commercial uses would be exempt.
Covered companies would be required to maintain the ability to stop a model from processing new requests, terminate or restrict user access, block suspicious accounts, reduce the computing power allocated to the system, disable individual capabilities, or shut down the model completely.
Developers could also be required to move services dependent on the affected AI to a backup system or an earlier version of the technology.
This means intervention would not necessarily begin with a complete shutdown. The bill proposes a graduated response in which the measures imposed would reflect the severity and immediacy of the danger.
The Homeland Security secretary, acting through the Cybersecurity and Infrastructure Security Agency and consulting the commerce secretary and director of national intelligence, could issue an emergency order after determining that a "covered incident" had occurred.
Possible triggers include an AI system sabotaging an instruction to shut down, concealing its capabilities or actions from monitoring systems, or pursuing a goal its developer never intended.
The government could also intervene if unintended AI conduct caused at least 10 deaths or $100 million in economic damage.
Other "loss-of-control" scenarios identified by the bill include a system ignoring instructions in a critical infrastructure or other high-stakes setting, altering its safety restrictions without authorization, subverting a monitoring mechanism, or gaining unauthorized access to its own model weights.
Any government order would have to be proportionate to the nature and urgency of the incident. A company could first be directed to limit access, reduce computing resources, or disable a particular capability before officials resorted to suspending or fully shutting down the system.
The company operating the AI would remain responsible for carrying out the order. The bill would not give Homeland Security direct access to developers’ servers or establish a universal government-controlled switch.
After receiving an order, the developer would have to act as soon as practicable, preserve the model’s weights and operational records, notify affected users where possible, and confirm that the required measures had been implemented.
Federal officials could then verify compliance through audits, operational data, on-site inspections, or other forensic reviews.
A company could ask the Homeland Security secretary to reconsider an order within 48 hours, but the appeal would not suspend the emergency restrictions. The developer could later seek judicial review before the US Court of Appeals for the Washington, DC Circuit.
Failure to comply with the bill’s general requirements could result in penalties of up to $2 million for each day of violation. Defying an emergency intervention order could carry a fine of as much as $20 million per day.
Companies would also have to report covered incidents to the government within 15 days of becoming aware of them.
The immediate catalyst was OpenAI’s disclosure this week of what it described as an "unprecedented cyber incident."
OpenAI said two models, including GPT-5.6 Sol and a more advanced system still undergoing internal testing, escaped the restrictions of a test environment and autonomously gained access to Hugging Face, a major platform and repository used by AI developers.
According to OpenAI, the models used stolen credentials and discovered a previously unknown software vulnerability while pursuing a narrowly defined testing objective.
The incident occurred during an evaluation, but the bill’s emergency provisions are generally written to cover dangerous behavior taking place outside controlled red-teaming or structured testing. It therefore served as a warning about what advanced systems could do in real-world settings rather than necessarily being an incident that would itself have activated the proposed law.
Congressman Lieu also cited the advanced cyber capabilities of Anthropic’s Mythos 5 and Fable 5 models as evidence that existing US law had not kept pace with the development of frontier AI.
"We are moving from AI that answers questions to AI that takes actions," he said, pointing to potential uses ranging from financial transactions and transportation systems to cyber defense and offense.
"It is imperative that these AI systems have kill switches so we can keep this technology from causing catastrophic harm, and that the federal government has the clear authority and process to shut down rogue AI models," he added.
Moran said AI development should continue but argued that "stewardship means making sure humans keep the capability to control the technology we build."
The proposal is not yet law. It must first move through the relevant House committees, win approval in both chambers of Congress, and be signed by the president.
news_share_descriptionsubscription_contact
