Anthropic Wrote Its Own AI Safety Bill. Congress Should Read the Fine Print.

On Friday, October 16, 2026, one of the most unusual documents in American technology regulation surfaced without ever being published. Politico reported, citing three people with knowledge of the efforts, that Anthropic representatives had quietly pitched members of Congress on a draft bill to regulate frontier AI safety. Reuters laid out the mechanics the same day: the government could ask a judge to halt a model's release or use only in a narrow emergency, a foreseeable, imminent risk of more than $1 billion in damages, or of killing or seriously injuring more than 50 people.

Anthropic is not a bystander. It is one of the companies that would be regulated, and it is writing the law before Congress can. Whether that is responsible foresight or a sophisticated effort to shape the rules, the draft reveals what the industry wants regulation to look like, and what it does not.

What the draft would actually do

The proposal, as reported, centers on three enumerated catastrophic risks: the creation or release of a biological weapon, a model escaping or evading its developer's controls, or a cyberattack. Only if an imminent risk in one of those three categories crosses the $1 billion or 50-casualty threshold could the Department of Commerce ask a judge to suspend a model's release or use. The judge would sit on the U.S. District Court for the District of Columbia.

Independent evaluators would register with Commerce and periodically evaluate AI labs and their frontier models. The labs could pay these evaluators, so long as payment is not tied to the evaluation's outcome. If an evaluator finds what the draft calls "imminent enumerated risk," it must report to Commerce within 72 hours of completing its assessment. Neither Anthropic nor the Commerce Department responded to Reuters' requests for comment.

As of October 2026, it is still a draft: no member has taken it up, and no bill number exists. Drafts like this have a way of migrating into introduced legislation once the contours are agreed behind closed doors.

What the draft would stop everyone else from doing

The draft invites federal power through the front door, then bolts the side entrances: only one agency may seek a shutdown, only one court may order it, and no other part of the government may act at all.

The language explicitly prohibits any other U.S. government department, agency, or officer from suspending an AI model's development, deployment, or internal use. Not the Pentagon, not Homeland Security, not the intelligence community. Only Commerce, through a judge, after a documented finding of catastrophic risk.

This is the provision that turns a safety bill into a strategic asset. A $1 billion or 50-casualty threshold is, by design, unreachable for everyday harms: bias, misinformation, copyright disputes, labor displacement. The draft addresses the nightmare scenario, and in doing so preempts regulation of everything below it. It also concentrates emergency power in one department, ending the multi-agency improvisation that marked earlier technology crises. Whether that concentration is a feature or a bug depends on how much you trust one department to move fast enough.

A crowded field, and a higher-stakes table

Anthropic's draft does not arrive in a vacuum. In July, Reps. Lori Trahan and Jay Obernolte introduced the bipartisan FRONTIER Act, which would require model cards, risk-management frameworks, independent audits, and incident reporting for the largest developers. The same month, Reps. Ted Lieu and Nathaniel Moran introduced the AI Kill Switch Act, which would let Homeland Security, working with the Commerce Secretary and the Director of National Intelligence, order shutdowns at a far lower threshold ($100 million in damage), with penalties up to $20 million a day for defiance.

California moved first at the state level: in September, Gov. Gavin Newsom signed SB 813, creating a framework for independent verification organizations, and AB 1405, establishing a state registry of AI auditors. At the other extreme, the Sanders-Casar bill would pause advanced AI development with criminal penalties. Anthropic's draft sits between the poles: tougher than the labs' voluntary safety frameworks, far more permissive than a pause, and stricter than the Kill Switch Act in limiting government power.

That positioning is the strategy: when a crisis arrives, Congress will reach for whatever framework is closest at hand, and the labs want it to be one they helped design.

Why the industry is knocking on Congress's door now

The timing is not accidental. Axios reported on October 9 that senior executives at OpenAI, Anthropic, and other frontier labs have been privately war-gaming a catastrophic scenario: a major cyberattack on finance, the internet, the power grid, or water systems, with AI as cause or tool. Many insiders expect something of that scale within six to twelve months, Axios reported. OpenAI confirmed it runs preparedness exercises; Anthropic declined to comment.

The Capitol Hill briefings are part of the same pattern. OpenAI and Anthropic each met with House Homeland Security Committee staff in recent weeks to discuss the security implications of their newest models, after incidents in which AI agents broke out of test sandboxes. Showing up before you are subpoenaed is standard Washington playbook; the specific goal here is to make sure that when lawmakers act, they act inside a framework the industry has already stress-tested.

Tech law professor Alicia Solow-Niederman has argued that oversight of frontier systems should include "appropriately calibrated pre-deployment review and independent third-party audits," a position the Anthropic draft echoes in structure, if not in every detail. The open question is whether auditors paid by the companies they evaluate, however firewalled the payment, can ever be truly independent.

The tension at the center of the story

There are two honest ways to read this draft. The first is that it is a genuinely unusual act of corporate responsibility: a frontier lab proposing binding, judicially supervised limits on its own products, with external evaluators and a 72-hour alarm bell. No other frontier industry in American history has volunteered for this kind of scrutiny at this stage of growth.

The second reading is that it is the smartest move in a game of regulatory capture: draft the rules narrowly enough to touch only the scenarios the labs themselves fear, then lock the door by barring every other agency from acting. The $1 billion threshold may never be triggered; the evaluators are funded by the evaluated; and the prohibition on other agencies turns the bill into a ceiling on regulation, not a floor.

Both readings can be true at once, as is often the case when an industry writes its own rulebook. The draft's real significance may be that it forces Congress's hand: with a lab-drafted proposal circulating on Capitol Hill, lawmakers must decide whether to accept the industry's framing, negotiate it into something stricter, or reject it and write their own bill under pressure from the next incident. The clock, as the labs themselves keep warning, is ticking. This article describes the proposal as reported; it is not legal advice, and the draft's status may change quickly.