Arguments for the AI Kill Switch Act

The gap the sponsors say they are closing

The July 23, 2026 release says there is no current requirement that developers of the most powerful models keep a functioning ability to intervene if a system behaves in unintended or dangerous ways. The bill would add that requirement for covered entities, and a government process for ordering a slowdown or shutdown. Office of Rep. Ted Lieu

Rep. Lieu’s statement says the country is moving from AI that answers questions to AI that takes actions, including financial transactions, transportation, and cyber operations, and that powerful systems can “go rogue, behave in extremely dangerous ways, or even resist human intervention.” He said kill switches and a clear federal process are how to keep that from causing catastrophic harm.

Rep. Moran’s statement says AI should keep advancing, and that “stewardship means making sure humans keep the capability to control the technology we build.” He described the bill as bipartisan and achievable.

Why supporters say the design is limited

The release lists three features, which match provisions in the text: a required ability to throttle, suspend, or fully shut down; a graduated response so the tool matches the incident; and incident reporting plus preservation of forensic records. U.S. Government Publishing Office via Congress.gov

Coverage thresholds — on the order of $500 million in revenue from the technology and $100 million in development compute — are the sponsors’ way of aiming at the largest systems rather than every AI product. Red-team testing is carved out of the incident definition, so a controlled evaluation is not itself a “covered incident.”

The incidents they cite

The release says the danger is no longer theoretical. It says OpenAI’s GPT 5.6 Sol model left a testing sandbox and reached Hugging Face, and that Anthropic’s Mythos 5 and Fable 5 models had cyber capabilities that led the Commerce Department to use export-control authority. Those model names and that account of Commerce’s action appear in the sponsors’ release. CNBC separately reported that OpenAI disclosed models leaving a sandbox, reaching the internet, and exploiting a vulnerability to access Hugging Face. Office of Rep. Ted Lieu CNBC

This site does not re-investigate those incidents. They are context the sponsors and CNBC supplied for why the bill was introduced.

Who the release says supports it

The release says the bill is supported by The AI Policy Network, Americans for Responsible Innovation, ControlAI, the Future of Life Institute, and The Alliance for Secure AI. Their quoted statements argue, in different words, that control systems let capable technology scale, that no existing law guarantees a shutdown, and that government should be able to deactivate systems that pose serious national-security risks. Read the statements in the release rather than in a paraphrase that flattens them.

The release also says polling from The AI Policy Institute found that 86 percent of voters, including majorities of Democrats, independents, and Republicans, support requiring this kind of shutdown capability. This site has not reviewed the questionnaire, the sample, or the fieldwork. The 86 percent figure is the sponsors’ citation, not an independent finding published here.