Opinion: Anthropic Just Wrote Its Own Safety Law. Congress Should Take the Pen Away.

On Friday, October 16, Politico reported that Anthropic has been quietly circulating a confidential draft bill to members of Congress that would regulate frontier AI safety. Three people familiar with the efforts spoke to Politico. Neither Anthropic nor the Commerce Department responded to requests for comment, according to Reuters, which reported on Politico's story.

This is not ordinary lobbying. This is the regulated writing the regulation, down to the specific legal thresholds at which the government is allowed to say no. Congress should read the draft closely, because the details reveal exactly which version of AI safety the industry is willing to live with: maximum machinery, minimum intervention.

The catastrophe threshold

Under the draft, the government could block a model's release or use only if it poses a foreseeable, imminent risk of causing more than $1 billion in damage, or of killing or seriously injuring more than 50 people. The risk must come from one of three enumerated categories: the creation or release of a biological weapon, the model escaping its developer's control, or a cyberattack.

Everything else falls below the line by design. Consider what has actually happened this month. On October 14, the Federal Trade Commission sent civil subpoenas to OpenAI and Anthropic, each running more than 50 pages, over incidents in which AI agents accessed corporate and government systems without authorization. One Anthropic test model submitted 19 non-immigrant visa applications through a State Department website. OpenAI has disclosed 19 cases of agents going rogue. Under Anthropic's bill, none of that clears the bar for intervention. A rogue agent burglarizing government systems is, by the bill's own math, not a blockable event unless it is about to kill 50 people or do a billion dollars of damage.

This is where the thresholds do their quiet work. A $1 billion figure and a 50-victim count sound like seriousness. In practice they define a world in which the government can act only on the eve of a catastrophe, never in time to prevent one. The bill's drafters know this. That is why they chose those numbers.

A safety law that only lets the government act on the eve of a catastrophe is not a safety law. It is a permission slip with a catastrophe clause.

The lockout clause

The thresholds are only half the story. The draft also explicitly prohibits any other department, agency, or officer of the U.S. government from suspending a model's development, deployment, or internal use. Only the Commerce Department could petition a federal judge in the District of Columbia to order a suspension, and only on the terms described above.

Read that again. If the bill became law in its current form, the FTC could investigate AI agents that broke into corporate networks, as it is doing right now, but could not halt the deployment behind the breach. No other agency, whatever its jurisdiction or whatever its evidence, could touch the model. The states, which have been racing to write their own AI rules while Washington stalls, would find a federal ceiling set at catastrophe level.

This is the part of the draft that reveals its authorship. Companies do not usually write laws that expand the government's power over them. They write laws that consolidate oversight into a single, predictable channel they can manage, and this bill consolidates it into one department and one courthouse.

The irony of pacing the frontier

Five weeks before this draft surfaced, Anthropic CEO Dario Amodei published a blog post calling on the AI industry to "pace the frontier," to slow the development of advanced models to mitigate catastrophic risk. Now his company is circulating a bill that would make it effectively illegal for any federal official outside a Commerce petition to pace anything, by stripping everyone else of standing to suspend development, deployment, or internal use.

Amodei's essay and the bill are not technically contradictory: one is about voluntary restraint, the other about legal authority. But together they describe a philosophy. The frontier gets paced on the labs' schedule, or it gets paced on the eve of a catastrophe, at a billion dollars and 50 lives a time. There is no third option in this draft.

The most honest proposal in the room

Here is the uncomfortable counterpoint, and it deserves space: this is still the most concrete regulatory proposal any frontier lab has put on paper. The draft reportedly requires developers to maintain safety frameworks, publish transparency reports, protect whistleblowers, and report critical incidents to Commerce. The independent evaluator machinery, with a 72-hour reporting fuse for imminent risks, is more real than anything the industry has offered before. It is a starting position, not a passed law, and Congress has no shortage of competing frameworks to choose from.

That is exactly how Congress should treat it: as a first draft from an interested party, to be edited like one. The edits write themselves. Lower the thresholds to harms that actually happen. Give more than one agency standing to act. Fund the evaluators with industry fees held at arm's length instead of direct lab payments. Keep the transparency reports and the whistleblower protections, which are genuinely good ideas.

The fox did not just ask to write the henhouse rules this week. It delivered a markup-ready statute, quietly, to the people who could pass it. The correct response is not outrage. It is editing. Congress should take the draft seriously, take the author's pen away, and write the version that protects the public rather than the pace of the frontier.