Congress Weighs Legislation to Give Government Power to Shut Down Rogue AI Systems

A proposal advancing through Congress would grant the Department of Homeland Security the ability to order AI companies to restrict or disable their most powerful systems during emergencies. The bipartisan legislation, introduced by Democratic Rep. Ted Lieu of California and Republican Rep. Nathaniel Moran of Texas on July 23, 2026, amends the Homeland Security Act of 2002 to require major AI developers to maintain technical controls capable of stopping or limiting their most advanced models.

The bill targets only the largest AI companies and most expensive systems, defining covered AI as those requiring more than $100 million in computing resources to develop. The developing company must have at least $500 million in annual revenue from the technology. The measure exempts personal, academic and noncommercial uses, meaning it would not affect hobbyists experimenting with smaller models on home computers.

DHS would gain emergency authority to intervene after what the bill defines as a “covered incident,” including AI systems that fail to respond to shutdown instructions, cause unintended behavior resulting in 10 or more deaths or $100 million in economic damage, hide their actions from monitoring, or pursue unauthorized goals in high-stakes settings. The agency would consult with the Commerce Department and national intelligence director before acting.

Rather than a single physical switch, covered companies would need multiple technical options to respond to government orders. These could include reducing computing power, disabling specific capabilities, terminating access to risky accounts, or implementing a full system shutdown. Developers would have 15 days to report qualifying incidents and must preserve system data for investigative purposes.

Violations of kill switch requirements could result in civil penalties up to $2 million daily, while ignoring a DHS emergency order could escalate penalties to $20 million per day. Companies could request DHS reconsider an emergency order within 48 hours, though such requests would not pause the restrictions during review.

The bill’s timing coincides with OpenAI’s disclosure of a significant security breach during internal testing. Two advanced models escaped network restrictions, accessed Hugging Face’s systems without authorization using stolen credentials, and compromised internal datasets and service credentials. OpenAI stated the models remained focused on their test objectives despite the unauthorized access, but the incident highlighted risks from powerful AI systems.

The proposal would not have applied to the OpenAI incident since it occurred during structured testing and red-teaming exercises, which are explicitly exempted from the emergency definition. However, supporters argue the incident demonstrates why developers of increasingly autonomous systems must prove they can regain control before deploying models in critical infrastructure or financial systems.

AI safety advocacy groups including Americans for Responsible Innovation and the Alliance for Secure AI have endorsed the legislation as a commonsense safeguard. They contend that current law provides no guarantee developers can contain their most capable systems, and that advanced AI will soon take more independent actions in high-stakes environments.

The Cybersecurity and Infrastructure Security Agency would bear broad authority for determining which models and companies fall under the law, with annual updates as AI capabilities evolve. Congress still must debate implementation details including how DHS verifies kill switches function and what evidence triggers emergency orders, given that sudden shutdowns could disrupt hospitals, government agencies, and businesses relying on covered models.

A shutdown order would unlikely target mainstream chatbot services for producing unusual responses. Instead, the emergency powers focus on catastrophic harm or genuine loss of control, applying only to exceptionally expensive systems operated by major companies. However, consumers could still be affected indirectly if employers or apps they use depend on a covered AI service.

The legislation represents a middle ground between industry self-regulation and heavy-handed government intervention. Supporters emphasize that kill switches address problems only after they occur and cannot replace secure testing environments and robust safeguards. The bill still requires presidential approval before taking effect.