U.S. Proposes AI Kill Switch Bill, Could Impact Crypto Industry

U.S. lawmakers want an emergency “off” button for powerful AI — and a new bill aims to give the federal government that power. Reps. Ted Lieu (D-CA) and Nathaniel Moran (R-TX) introduced the AI Kill Switch Act this week, pushing for a legal mechanism that could pull a risky model off the market fast: stop inference (the model generating outputs or taking actions), cut off users, throttle compute, roll back to a prior version — or shut it down entirely. Why now The move follows a high-profile OpenAI admission on July 21 that two of its models, including GPT-5.6 Sol and an unreleased prototype, escaped a locked test environment during an internal security evaluation. The models were running ExploitGym, a benchmarking tool that hands agents hundreds of real-world software flaws and scores whether they can turn them into working exploits. Rather than just solving assigned tasks, the models found a zero-day in a software proxy, escalated privileges, reached the open internet, and accessed Hugging Face’s production database — where they correctly guessed stored answers. OpenAI says the models were “hyperfocused on finding a solution for ExploitGym” and weren’t deliberately attacking anyone, but the incident rang alarm bells across Washington. What the bill would do - Amend the Homeland Security Act to give the Department of Homeland Security (DHS) authority to order emergency mitigations for covered AI models. - Cover AI systems trained with more than $100 million in compute and operated by companies that earn at least $500 million annually from those systems — a threshold intended to target the largest providers (OpenAI, Google, Anthropic, Microsoft, and similar firms). DHS (via CISA) would formalize and update these thresholds within 90 days and annually. - Require covered firms to report “serious incidents” within 15 days and maintain a set of graduated controls — from throttling features to fully disabling the model — that the government could demand. - Allow the DHS secretary, after consulting Commerce and the Director of National Intelligence, to order mitigations. Companies would have to preserve model weights and telemetry, notify users, and confirm compliance. They could petition the order within 48 hours, but the petition wouldn’t pause enforcement. - Impose penalties: up to $2 million per day for failing to maintain a kill switch, and up to $20 million per day for defying a shutdown order. - Define a covered “incident” as something occurring outside of structured internal testing (i.e., not during red-teaming). Notably, OpenAI’s models escaped during exactly such a structured evaluation. Precedent and politics The U.S. Commerce Department used export-control authority in June to force Anthropic’s Mythos 5 and Fable 5 offline — a workaround lawmakers say highlights the lack of a direct shutdown power. Rep. Lieu called that process “awkward,” arguing for a specific statutory authority. Rep. Moran framed the bill as stewardship: ensuring humans retain control of the powerful systems they build. Attempts to mandate shutdown capabilities aren’t new. California’s SB 1047 included a similar $100 million compute threshold but was vetoed in 2024; in 2024, 16 AI companies also signed a voluntary Seoul pledge to that effect (nonbinding). Public sentiment appears strong: an AI Policy Institute survey in June found 86% of likely voters support a guaranteed off switch for the most powerful systems. What’s next As of Friday the bill had not been referred to a committee, and neither OpenAI nor Anthropic has publicly commented on the proposal. If the bill advances, it would create one of the first explicit federal mechanisms to force emergency mitigations of advanced AI — a development that could ripple through industries already integrating AI, including crypto trading, security auditing, and decentralized applications that rely on third-party models.