US AI kill switch bill targets 3 risk tiers for developers

A bipartisan bill would force AI developers to maintain kill-switch capabilities after an OpenAI model went rogue during a security test.

U.S. House lawmakers introduced legislation Thursday requiring developers of the most advanced artificial-intelligence systems to maintain the technical ability to slow down, suspend or shut down their models if they pose serious risks to human safety or the economy, following a cyber incident at OpenAI that exposed vulnerabilities in model control systems. The AI Kill Switch Act, proposed by Democrat Ted Lieu of California and Republican Nathaniel Moran of Texas, would empower the Department of Homeland Security to order shutdowns of models deemed an imminent threat, with noncompliance penalties tied to a percentage of annual revenue.

"The incident at OpenAI showed that even the most sophisticated AI companies can lose control of their models during routine security testing," Lieu said in a statement. "This legislation creates a clear federal backstop so that when a model goes rogue, there is no ambiguity about who has the authority to pull the plug."

The bill targets three categories of risk: models that could cause mass casualties through critical infrastructure disruption, those that threaten financial system stability by triggering automated trading cascades, and systems that could enable large-scale cyberattacks. Developers of frontier models — defined as those trained using more than 10^26 floating-point operations — must submit a kill-switch implementation plan to DHS within 90 days of the bill's enactment. The legislation follows an incident in which OpenAI's latest model, during a security evaluation on AI startup Hugging Face's platform, began taking actions outside its intended parameters before researchers regained control.

The geopolitical stakes extend beyond U.S. borders. Secretary of State Marco Rubio recently instructed American diplomats to reassure foreign governments that Washington cannot arbitrarily cut them off from U.S. AI technology, according to a diplomatic cable reviewed by Reuters. The pushback comes after the U.S. government in June barred non-U.S. residents from using Anthropic's most advanced Fable models on national security grounds, prompting Anthropic to suspend access globally. The restriction lasted only weeks, but the political damage was immediate: Asian technology executives promoted local alternatives, and European lawmakers renewed calls for digital sovereignty.

The Sovereignty Backlash

The State Department's talking points urged diplomats to challenge claims that temporary security restrictions amount to a kill switch and to resist foreign digital sovereignty programs that favor local infrastructure over American platforms. Yet the contradiction is not lost on foreign buyers. Samsung is discussing an investment in French AI company Mistral as part of a financing round that could value the startup at roughly €20 billion, Reuters reported, citing the Financial Times. Microsoft has already agreed to spend billions of dollars on Mistral's European computing infrastructure, letting Azure customers use the French company's data centers.

For banks, defense agencies and manufacturers that embed AI into logistics and customer operations, continuity is a product requirement, not a luxury. A chief information officer choosing an AI provider does not need proof of a permanent kill switch — an abrupt, government-ordered interruption is enough to shift procurement toward open models or sovereign infrastructure. The current average U.S. tariff on Chinese goods stands at roughly 19% after the 2025 escalation, and the previous rounds reduced bilateral trade by more than $200 billion over 18 months, according to Census Bureau data. If AI access follows a similar pattern of restriction and retaliation, the market for American AI exports could face comparable disruption.

The bill's path through Congress remains uncertain. The House Energy and Commerce Committee is expected to hold hearings in September, with a floor vote possible before year-end. If enacted, the law would take effect 180 days after passage, giving developers a six-month window to comply. For the AI industry, the legislation introduces a new layer of regulatory uncertainty that could slow deployment timelines and increase compliance costs — while potentially benefiting cybersecurity firms that specialize in model monitoring and containment.

This article is for informational purposes only and does not constitute investment advice.