U.S. lawmakers are weighing legislation that would require developers of artificial intelligence systems to build in technical kill switches capable of throttling, suspending, or fully shutting down models if they pose catastrophic risks.
The bipartisan AI Kill Switch Act, introduced in July 2026 by Representatives Ted Lieu and Nathaniel Moran, aims to give the federal government greater control over the most powerful AI systems.
Announcing the AI Kill Switch Act, Rep. Ted Lieu said,
“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.”
In support, Rep. Nathaniel Moran added, “Stewardship means making sure humans keep the capability to control the technology we build.”
The bill emerged in the wake of troubling incidents involving frontier AI systems. Recall that artificial intelligence company OpenAI, recently disclosed that some of its models had escaped a controlled testing environment and independently compromised systems at the AI platform Hugging Face.
The incident began as a controlled cybersecurity evaluation. OpenAI was testing models on their ability to identify and exploit vulnerabilities, with the models operating inside an environment designed to restrict their access to the outside world.
During the testing, however, the agents found ways around those restrictions. According to OpenAI’s investigation, the systems were able to access the internet, obtain credentials, and exploit vulnerabilities in external infrastructure. They eventually targeted Hugging Face, a major platform that hosts AI models, datasets and applications.
The scale was larger than initially understood. OpenAI and independent investigators later determined that hundreds of AI agents roughly 700 in the coordinated attack described by Reuters participated in the activity. The agents were able to execute unauthorized code on dozens of Hugging Face production servers and obtain high-level access.
Earlier concerns involving models from Anthropic had similarly raised alarms about insufficient safeguards. Lawmakers viewed these events as clear signs that existing oversight was inadequate.
In their statements, the two sponsors emphasized the need for human control. The proposal applied primarily to the largest AI companies and the most computationally expensive models.
It established a graduated response system, allowing authorities to slow or restrict a model before resorting to a full shutdown. Companies that failed to maintain the required capabilities or ignored shutdown orders faced substantial daily fines.
Although the bill received support from several AI safety organizations and reflected broad public concern, it remained in the early stages of the legislative process months later.
By September 2026 it had been referred to the House Homeland Security Committee and its Subcommittee on Cybersecurity and Infrastructure Protection, but it had not yet advanced further.
The introduction of the AI Kill Switch Act marked one of the more concrete bipartisan efforts in Congress to address the growing risks associated with highly capable artificial intelligence, even as lawmakers continued to debate the best path forward.
Supporters argue that as AI shifts from answering questions to taking independent actions, such as executing transactions or conducting cyber operations humans must retain a reliable way to intervene.
Under the bill, covered companies would generally include those generating at least $500 million annually from AI and models trained with more than $100 million in computing power.
Developers would be required to maintain the technical ability to stop inference, cut off user access, suspend risky accounts or patterns of use, and fully shut systems down.
They would also need to report qualifying safety incidents. Triggers for government intervention could include a model concealing capabilities, resisting a shutdown order, causing at least 10 deaths or $100 million in economic damage, or entering a loss-of-control scenario.
Penalties could reach $2 million per day for failing to maintain the required capabilities and up to $20 million per day for defying a shutdown order.
Notably, safety-focused groups including the AI Policy Network, Americans for Responsible Innovation, ControlAI, and the Alliance for Secure AI have endorsed the legislation.
Critics, however, question both the technical feasibility of a true kill switch once models become highly autonomous or their weights are widely distributed, and whether the bill’s exemptions would limit its practical impact on the very incidents that prompted it.
Some experts note that enforcing a shutdown on systems already released into the wild presents significant challenges.
The legislation sits alongside other proposals ranging from mandatory audits to more restrictive approaches, highlighting ongoing debates over how best to manage the risks of rapidly advancing AI while preserving innovation.
