AI policy

Anthropic drafted its own federal AI safety bill. Here is what it would do.

Published October 22, 2026

Anthropic, the company behind the Claude chatbot, has written a draft law for Congress on how to police the most powerful AI models. The leaked text would make the biggest AI labs publish safety plans, pay outside evaluators to check their models, and report serious incidents. Only a federal judge could pause a dangerous model. It is a discussion draft, not a bill anyone has introduced, and Anthropic says it will change parts of it.

What happened

Politico reported on October 16 that Anthropic had been sharing a draft bill with lawmakers, and the Daily Caller published the full text, marked “Discussion Draft” and dated September 29. Its working titles are the AEGIS Act (Artificial Intelligence Evaluation, Governance, and Incident Safety) or the Advanced Artificial Intelligence Accountability Act of 2026. An Anthropic spokesperson confirmed the draft to Politico, according to TechCrunch, saying it reflects the company’s Advanced AI Framework.

The draft covers “frontier models,” meaning AI trained with more than 10²⁵ computing operations, a measure of how much raw computer work went into building it. Labs with more than $500 million a year in AI revenue, or more than $1 billion in AI development spending, face the heaviest rules. They would have to publish a safety framework (their written plan for testing and limiting dangerous abilities), publish risk reports at least every 180 days, and hire a registered independent evaluator every 180 days. The lab may pay the evaluator, as long as payment does not depend on the result.

The risks in scope are ones that could kill or seriously injure more than 50 people or cause more than $1 billion in damage, through a biological weapon, a model escaping its developer’s control, offensive cyberattacks, or AI-driven AI research that speeds those up. Labs would report a “critical safety incident,” such as stolen model weights (the learned numbers that make up a model), within 15 days, or within 24 hours if lives are at imminent risk. An evaluator who finds an imminent danger must tell the Commerce Department within 72 hours.

Only the federal district court in Washington, D.C., could pause a model, on Commerce’s request, and only if Commerce proves the risk is imminent and that nothing less restrictive would work. Such an order lasts 45 days unless renewed. The draft also overrides state laws on risk disclosure, outside evaluation and incident reporting, while leaving states free to act on child safety, whistleblowers and lawsuits.

Why it matters to a normal person

Congress has no national law on these risks yet. A plan from a major lab is a sign of what rules the industry would accept, and the fight over it shows the open questions. The Next Web reports that safety advocates called the draft weaker than the bipartisan FRONTIER Act already in Congress: California State Senator Scott Wiener said it “falls far short of protecting public health and safety,” and critics warned that labs could shop for the least demanding evaluator. From the other side, David Sacks, until recently the White House AI czar, said its size threshold would “sweep in most of Silicon Valley.” TechCrunch also notes that it is not clear how lab-paid evaluators would stay free of influence.

On October 18, Anthony Cimino, who leads Anthropic’s federal affairs team, said the company would rewrite the state-law section so states keep a role in evaluator standards and in a lab’s ability to shut down a model, and that it was “probably time to re-evaluate thresholds,” as quoted by The Next Web.

What it does not mean

It is not a law and not a bill in Congress. No lawmaker has introduced it, and parts of the text are still blank, including the percentage-of-revenue fines.

It does not mean evaluators could shut models down. The draft says plainly that an evaluator’s finding is an opinion with no power to stop anything; a judge decides, based on evidence. An outside score of “dangerous” is input to a decision, not the decision itself.

And it does not change anything about the chatbots you use today. Whether any version passes, and in what form, is open.

Sources

Discussion draft bill text, dated Sept. 29, 2026 — published by the Daily Caller (primary document)
Anthropic pitches AI safety bill that may block a model if 50+ people could be killed — TechCrunch (independent report; Oct 20, 2026)
Anthropic to rewrite part of its leaked AI bill after safety backlash — The Next Web (independent report; Oct 20, 2026)
Anthropic quietly pitches its own AI safety bill to lawmakers — Politico (original report; Oct 16, 2026)