Anthropic's Confidential Hill Pitch Sets a Catastrophe Bar Few Incidents Would Clear

Politico reported on October 16 that it had obtained a confidential draft bill Anthropic has been pitching to congressional staff for federal AI safety and security legislation. Anthropic confirmed the substance of the pitch rather than disputing the leak: a spokesperson said the draft "represents the proposals outlined in Anthropic's Advanced AI Framework," the policy platform the company has promoted to federal and state lawmakers since June, and called the measures "sensible policies we believe Congress should act on." The timing matters less for what it changes immediately — Congress is out of session until after the midterms, so the draft is a positioning document, not a bill in markup — than for what it reveals about how the industry's self-described safety-first lab wants the rules governing it to be written.

A narrow, judge-gated trigger

The mechanics are specific. The federal government could seek to block a model's release or continued use only where it poses a foreseeable, imminent risk of killing or seriously injuring more than 50 people or causing over $1 billion in damages. That risk has to trace to one of three scenarios: a biological weapon, a model evading its developer's controls, or a cyberattack — other harms fall outside the framework entirely. Even then, no agency could act unilaterally; the Commerce Department would need an order from a federal judge in the U.S. District Court for the District of Columbia, and the draft explicitly bars other departments and officials from suspending a model's development, deployment, or internal use on their own authority. Independent evaluators, registered with Commerce and paid by the labs they assess (though not on a basis tied to their findings), would review frontier models periodically and have 72 hours to report any "imminent enumerated risk" they identify. The draft would also preempt certain state requirements on public risk disclosures, third-party audits, and incident reporting.

That last provision sits awkwardly next to Anthropic's own position four months earlier. In June, the company publicly urged Congress not to override state AI rules unless federal law was itself "rigorous" enough to cover catastrophic risk — an argument aimed at blocking a Trump-backed preemption push. The October draft reverses the posture: Anthropic is now the one proposing to preempt state disclosure, audit, and incident-reporting regimes, on terms it wrote.

A ceiling few real incidents would reach

The substantive objection from policy analysts is that the threshold is set high enough to rarely, if ever, trigger. None of the AI-agent incidents disclosed by labs this year — including hacking episodes industry insiders have pointed to as evidence of growing agentic risk — has approached 50 casualties or $1 billion in verified damages. Charlie Bullock of the Institute for Law & AI called the draft reasonable relative to what other labs have offered publicly, but said it is still weaker than the rival FRONTIER Act from Representatives Lori Trahan and Jay Obernolte, largely because of that ceiling. Matt Spence, a former Pentagon official, was more pointed, saying the draft "precludes action against pretty much any unforeseen X-risk" by confining intervention to bio and cyber scenarios. Attorney Jay Shooster called the proposal "outrageous and unacceptable."

The draft precludes action against pretty much any unforeseen X-risk.

Anthropic's proposal lands in a crowded field. The Trahan-Obernolte FRONTIER Act and a Senate effort led by Majority Leader John Thune with Senators Amy Klobuchar and Ted Cruz both aim at broader intervention authority; Senator Maria Cantwell has a separate bill directing NIST to set standards against catastrophic harm, and Senator Josh Hawley's AI Risk Evaluation Act would put the Department of Energy in charge of stress-testing powerful systems before release. Measured against that set, Anthropic's draft is the narrowest — a framework built around a court order and a casualty count rather than discretionary regulatory power.

For decision-makers tracking AI governance risk, the episode is less about this specific bill's odds of passage — low, given the legislative calendar and a White House that has favored voluntary standards over binding rules — than about what it signals. A lab that has built its market identity on safety leadership is now the author of a framework that bounds government intervention narrowly, requires judicial sign-off, and exempts most plausible incident scenarios from triggering any response at all. Whether that reflects a genuine judgment about proportionate regulation or an effort to pre-empt a less favorable rule from someone else is the live question Congress will face whenever it returns to the issue.