What happens when the company being regulated drafts the regulations? Anthropic is quietly shopping its own federal AI safety bill on Capitol Hill, and the government could only pull the plug on a model after a $1 billion or 50-death threshold shows up. That is the headline detail in the draft, according to Reuters’ account of a Politico report. In effect, the company is writing the rules it may someday live under.
The draft is confidential and has not been formally introduced. So far, no sponsor or bill number has surfaced. Politico reported the pitch on Friday, Oct. 16, citing three people with knowledge of the effort. We could not review the draft text, which is not public.
What the Anthropic AI safety bill would let Washington do
The core of the proposal is a narrow off switch, and a very specific one. The government could block a model’s release or use 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.
The risk would also have to come from one of three sources: a biological weapon, a model slipping its developer’s controls, or a cyberattack. Everything else falls outside that power.
Here is the part that matters most. Only the Commerce Department could act, and it could not do so alone. It would have to ask a federal judge in the District of Columbia to order a suspension, Business Today reports.
The language also bars any other U.S. agency or officer from suspending a model’s development, deployment or internal use. That is why Tom’s Hardware says the bill would effectively make Commerce the only agency able to stop an AI model. One button, one keyholder, one judge to sign off.
Testing and fine print in the Anthropic draft AI legislation
The Anthropic AI safety bill is not only a brake. Independent evaluators would register with Commerce and periodically test the lab and its frontier models. Labs could pay them, as long as payment is not tied to the outcome of the evaluation.
If evaluators spot an imminent covered risk, they would have 72 hours after finishing their assessment to report it to Commerce. Reporting also says developers would need safety frameworks, transparency reports for new models, whistleblower protections and critical-incident reporting. That echoes Anthropic’s own warning about AI systems copying themselves onto everyday devices, the kind of loss-of-control risk the bill targets.
One more detail is worth watching. The draft would reportedly override state and local rules on public risk disclosures, third-party audits and incident reporting. In plain terms, one federal rulebook would replace a growing patchwork.
Who sets the bar in this frontier AI safety bill?
So who benefits from a $1 billion bar? Critics argue the thresholds leave most AI harms outside federal halt authority. AI-powered fraud, for one, reportedly falls outside the covered risk categories. Limiting the power to Commerce could also complicate a response to risks that cross several agencies.
Still, the numbers are not new. They echo New York’s Responsible AI Safety and Education Act, which defines catastrophic risk as a foreseeable risk of death or serious injury to more than 50 people, or more than $1 billion in property damage. The risk categories reportedly track California’s AI laws, including SB 53.
Here’s the thing: this fight is about who gets to hit the brakes on AI, and how bad things must get first. That matters to anyone whose work, apps or personal data run through these models.
Anthropic’s pitch is that this is the sensible middle. A spokesperson told Politico the draft reflects its Advanced AI Framework, which the company has pushed with federal and state lawmakers since June. “These are sensible policies we believe Congress should act on,” the spokesperson said. Meanwhile, the White House has leaned the other way, with Trump backing a full-speed push in the AI race.
Why the Anthropic federal AI proposal lands now
The pitch lands while regulators are already circling. The FTC recently sent civil subpoenas to OpenAI and Anthropic over AI safety incidents and possible consumer-protection violations, Engadget reports. An agency spokesperson confirmed it.
Notice the contrast. Under Anthropic’s draft, a regulator could not halt a model over those kinds of harms. Under the FTC’s approach, it can at least start asking questions.
Reuters said Anthropic and Commerce did not immediately respond to its requests for comment, and no congressional reaction has surfaced. Congress is weighing competing frontier-AI proposals, and this one is described as the narrower approach. For more on where the administration stands, see how Trump renamed AI and named a task force.
Next up: whether any lawmaker puts a name on it. Until then, the most interesting Anthropic AI safety bill in Washington is one nobody has officially introduced.
