Anthropic Is Quietly Pitching Congress Its Own AI Safety Bill — With a $1 Billion Tripwire

WASHINGTON · SUNDAY, OCTOBER 18, 2026

The company that built its brand on AI safety has written its own AI safety law — and the number at the center of it is $1 billion. Politico reported on October 16 that Anthropic has quietly pitched Capitol Hill lawmakers on a draft bill that would let the federal government block a frontier AI model only if it poses a foreseeable, imminent risk of more than $1 billion in damages, or of killing or seriously injuring more than 50 people. Politico obtained the confidential draft and spoke to three people familiar with the effort, in a report by Owen Dahlkamp. An Anthropic spokesperson told Politico: "These are sensible policies we believe Congress should act on," adding, "This bill represents the proposals outlined in Anthropic's Advanced AI Framework" — the framework the company published in June 2026. Bloomberg also reported that Anthropic presented its own draft bill to lawmakers.

Pause on that threshold for a second, because everything else in this story flows from it. Under Anthropic's draft, a model that foreseeably risks $900 million in damage could not be blocked. A model that risks seriously injuring 40 people could not be blocked. The government's emergency brake exists, but it is bolted to the floor at catastrophe level — and only three kinds of catastrophe count. According to Politico, the risk must fall into one of three categories: the model helping create a biological weapon, a cyberattack, or the model evading its developer's controls. Fraud at scale, election manipulation, autonomous scams, deepfake abuse — none of those, however large, would meet the draft's definition of a blockable risk, because they are not on the list. The draft's language tracks California's SB 53, the state frontier-AI law, which uses similar catastrophic-harm framing.

How the machinery would work

The enforcement design is as deliberate as the threshold. According to Politico, independent evaluators would register with the Department of Commerce and periodically assess frontier labs and their models. The labs could pay those evaluators — but payment could not be tied to the evaluation's outcome, a safeguard meant to keep the money from buying the verdict. If evaluators find an "imminent enumerated risk," they must report it to Commerce within 72 hours of completing their assessment. Then comes the narrowest part of the whole design: only the Commerce Department could act on such a report, and even Commerce could not suspend a model itself. It would have to ask a federal judge for the District of Columbia, who alone could order a suspension. The draft explicitly bars every other federal department, agency or officer from suspending a model's development, deployment or internal use. Reuters, summarizing Politico's reporting, likewise noted that while Commerce could request a judge to halt a model, the language prohibits any other part of the government from doing so.

Around that core, the draft reads more like a full regulatory program. According to Politico, it would require frontier developers to publish safety frameworks, issue transparency reports, protect whistleblowers and report incidents to Commerce. It would also preempt some state AI laws — federal rules would override state requirements on risk disclosure, outside evaluation and incident reporting. That preemption piece matters enormously in practice: California and other states have been writing the only binding frontier-AI rules in the country, and a federal law that displaces them while setting a $1 billion federal tripwire would, critics argue, lower the total amount of oversight even as it creates the appearance of a national standard. As of October 17, the draft had been privately circulated and had not been formally introduced in Congress — this is a pitch, not yet a bill with a number.

The critics noticed the gap — including in Anthropic's own framework

The pushback arrived fast, and the sharpest version of it is not "this is too weak" in the abstract. It is, as critics said, weaker than what Anthropic itself proposed in June.Mother Jones, in its analysis of the draft, called the threshold alarming and pointed out that a model that might kill or seriously injure only a few people, or cause less than a billion dollars in damage, would sit beyond the government's blocking power. Zvi Mowshowitz, in an October 17 analysis of the draft, argued that its preemption of state law is stronger than the bipartisan FRONTIER Act sponsored by Representatives Trahan and Obernolte, while its enforcement and auditor-licensing provisions are weaker than Anthropic's own June framework. Representative Alex Bores, one of the House's most engaged members on AI safety, likewise called the draft weaker than Anthropic's framework, and Nathan Calvin argued it is less protective than existing law. When the people making that argument include lawmakers and analysts who generally want a federal AI safety law, the draft has a persuasion problem that money and access alone may not fix.

To be fair to Anthropic — and the fairness matters, because the alternative on the table is often nothing — the company would say, and its spokesperson did say, that these are the proposals from its Advanced AI Framework, offered as sensible policy Congress should act on. There is a real argument that a clear, high, judge-checked threshold beats a vague standard that lets any agency halt any model on a hunch, chilling development while China races ahead. The 72-hour reporting duty, the registered independent evaluators, the whistleblower protections and the transparency reports are not nothing; they are the skeleton of an oversight system. The fight is over where the skeleton's spine sits: whether "catastrophic" should be defined so high that almost nothing that has actually gone wrong with deployed AI so far would qualify, and whether states should be cleared out of the field in exchange.

A crowded Hill, and a company writing its own rules

Anthropic's draft does not land in a vacuum. In the Senate, Majority Leader John Thune and Senators Amy Klobuchar and Ted Cruz have been negotiating a catastrophic-risk bill of their own, and any final legislation will have to reconcile that effort with the House's FRONTIER Act and with Anthropic's thresholds. From the other direction, Representative Greg Casar and Senator Bernie Sanders have called for halting advanced AI development altogether — a position that makes Anthropic's draft look, by comparison, like the industry's moderate wing. That spectrum is exactly why the draft matters: in a negotiation that runs from "halt everything" to "regulate nothing," the text that a major lab quietly circulates can become the de facto center of gravity, especially when it arrives fully drafted while other bills are still being negotiated.

There is also a simple optics problem that no amount of framework citations will dissolve. A company asking Congress to adopt, as law, the standard the company wrote for itself — with a blocking threshold set at a round, enormous number, a single friendly agency, a single court, and preemption of the state laws that currently bite hardest — will be read by skeptics as regulatory capture with good graphic design. Anthropic's counter is that it published the framework publicly in June and has advocated for it with federal and state lawmakers since, which is true and also not quite an answer: publishing your preferred law before you pitch it is transparency about the pitch, not neutrality in it. How lawmakers weigh that will decide whether this draft becomes a bill, a bargaining chip, or a cautionary tale. It is the second major AI-business story on our pages today, after this morning's look at the debt fueling the AI boom — read both and you get the industry's 2026 in miniature: enormous money, enormous models, and a rulebook still being written by the people it will govern. More coverage in the news section and our games desk.