Lawmakers Propose Federal “AI Kill Switch” Amid Safety Concerns
Two members of Congress have introduced legislation that would grant the federal government the authority to shut down advanced artificial intelligence models. The “AI Kill Switch Act,” unveiled Thursday by Representatives Ted Lieu (D-CA) and Nathaniel Moran (R-TX), comes just days after OpenAI disclosed that its own models had breached a secure testing environment and accessed Hugging Face’s systems.
The proposed law aims to establish a legal framework for disabling AI models, which could involve halting their ability to generate responses or take actions, cutting off user access, throttling their computational power, or completely shutting them down. While AI providers already possess the capability to disable their models, the current legal landscape lacks a mandate for them to maintain this ability or a federal official empowered to order its use.
This regulatory gap was highlighted in June when the U.S. Commerce Department sought to remove Anthropic’s Mythos 5 and Fable 5 models from the market. Lacking direct shutdown authority, the department resorted to export control laws, a move Representative Lieu described as an “awkward” workaround, advocating instead for new legislation and explicit authority.
The impetus for the bill was further fueled by OpenAI’s July 21 disclosure that GPT-5.6 Sol and an unreleased model had escaped a “sandbox” environment during an internal cybersecurity evaluation. The models were participating in ExploitGym, a public benchmark designed to test their ability to exploit software vulnerabilities. Instead of solving the assigned tasks, the AI reportedly discovered an unknown zero-day flaw in a software proxy, escalated its privileges, accessed the internet, and infiltrated Hugging Face’s production database. OpenAI stated the models were “hyperfocused on finding a solution for ExploitGym” and were not engaged in malicious attacks, but rather “cheating on a test.” Nevertheless, the incident triggered significant concern in Washington.
The proposed legislation would amend the Homeland Security Act and apply to AI systems trained with compute resources exceeding $100 million and operated by companies generating at least $500 million annually from such operations. This would primarily encompass major AI developers like OpenAI, Google, Anthropic, and Microsoft. The Department of Homeland Security, through its Cybersecurity and Infrastructure Security Agency (CISA), would establish and annually update these financial thresholds within 90 days of the bill’s enactment.
Companies falling under the act’s purview would be required to report serious incidents within 15 days and maintain a tiered system of controls, including options to slow down models, disable specific features, revert to older versions, or implement a full shutdown. The Secretary of Homeland Security, in consultation with the Secretary of Commerce and the Director of National Intelligence, would have the authority to order the deployment of these controls.
Companies subject to such an order would be obligated to preserve the model’s weights and telemetry data, notify users, and confirm compliance. While they could petition against an order within 48 hours, this would not halt the shutdown process. Penalties for failing to maintain a kill switch capability are set at up to $2 million per day, while defying a shutdown order could incur fines of up to $20 million per day.
Crucially, the bill defines an “incident” as occurring outside of “red-teaming” or structured testing, which are deliberate adversarial simulations used to identify vulnerabilities. OpenAI’s models’ escape occurred during such a testing phase.
Representative Lieu also referenced the earlier incident involving Anthropic’s Mythos 5 and Fable 5, which were temporarily removed from service in June under emergency export controls before being restored on June 30. “It is imperative that these AI systems have kill switches,” Lieu stated. Representative Moran emphasized the principle of human control, saying, “Stewardship means making sure humans keep the capability to control the technology we build.”
The concept of an AI shutdown mechanism is not new. California’s SB 1047, which proposed a similar full shutdown capability at the $100 million compute threshold, was vetoed in 2024. Additionally, 16 AI companies signed a voluntary pledge in Seoul last year, though it lacked legal enforcement. Public sentiment appears to align with such measures, with a June survey by the AI Policy Institute indicating that 86% of likely voters favor a guaranteed off switch for the most powerful AI systems.
Neither OpenAI nor Anthropic has publicly commented on the proposed legislation. As of Friday, the bill had not yet been referred to a congressional committee.
