After a year of talk about pauses, moratoriums and open letters, Washington has produced something with actual teeth, or at least the draft of one. Reuters reported last week that Senate Majority Leader John Thune, Commerce Committee chairman Ted Cruz and Democrat Amy Klobuchar are drafting a bill that would put a binding legal "duty of care" on the makers of the most powerful AI models, and hand the government the power to block a model's release if it is judged unsafe. It is the most concrete move yet to turn the industry's voluntary safety promises into law, and the trio's combination of votes and committee control means it might actually move.
The phrase "duty of care" is doing the heavy lifting, and it matters more than it sounds. In law, a duty of care obliges you to take reasonable precautions against harms you can foresee. Applied to AI developers, it would mean companies are not just asked to publish safety frameworks or submit to audits, they would be legally required to design their systems to prevent catastrophic outcomes, and could be sued if they fail. Compare that with the House's FRONTIER Act from July, which demands audits and incident reports backed by fines. That is punishment after something breaks. The Senate approach creates liability for the design choice itself, with courts, not just regulators, deciding what "reasonable" looked like.
The bill adds a second, sharper tool: a reserved right for the government to stop a model from shipping, with companies free to challenge the decision in federal court. That power is not entirely new in spirit. The Trump administration has already twice used export-control law to restrict frontier models, once suspending Anthropic's Mythos 5 and Fable 5 worldwide, once limiting OpenAI's release of a model to approved partners. But those were trade-law improvisations. This would be a purpose-built authority, with an appeals process, aimed squarely at AI.
What lit the fuse was a rough 72 hours. A researcher's viral resignation from Anthropic, warning that the labs are "gambling with our lives," was followed by the company's own threat report conceding that its newer models can no longer be assumed to fall below the threshold for helping someone develop biological weapons. "This is no longer a certainty," Anthropic wrote, the first time a major lab has said so in public. That admission maps precisely onto the bill's stated purpose of heading off biological and nuclear catastrophes, and gave the drafters an uncomfortable real-world example to point at.
The politics, predictably, are a maze. President Trump has dismissed existential worries as things "that won't happen" and frames the whole subject as a race with China he does not want to lose. His allies go further: David Sacks, who advises the White House on technology, has suggested the safety warnings are a "doomer" influence operation, possibly financed by Beijing, to kneecap American innovation. On the other side, Bernie Sanders wants to ban superintelligence outright, and Josh Hawley has opened an investigation into OpenAI. Even a bill with three well-placed sponsors has to thread between a president who thinks it is unnecessary and colleagues who think it is too timid. A contested preemption clause, which would stop states from enforcing their own catastrophic-risk laws, echoes the state moratorium the Senate rejected 99 to 1 last year.
The largest obstacle, though, may simply be the calendar. With the November midterms bearing down, the Senate has roughly three weeks of session left and the House barely one. Capitol Hill sources told Semafor this is "the only viable option" before 2027, because a new Congress in January will bring different priorities and a different mood. So the most serious AI safety bill yet written faces a deadline measured in days. Whether it reaches a floor vote is genuinely uncertain. What is clear is that the era of the industry marking its own homework is, for the first time, under real legislative threat.