One way to deal with regulation is to wait for it. Another is to write it yourself. Anthropic has been quietly showing congressional offices a draft federal bill on frontier AI safety and security, according to three people familiar with the effort, as Owen Dahlkamp reported for POLITICO, which obtained the confidential text. The company does not deny it. “This bill represents the proposals outlined in Anthropic’s Advanced AI Framework, which we have been advocating for with federal and state lawmakers since June,” a spokesperson said. “These are sensible policies we believe Congress should act on.”
The draft has a name, the Artificial Intelligence Evaluation, Governance, and Incident Safety (AEGIS) Act. According to the Daily Caller, it is dated September 29, has gone to a small number of offices and, as of Monday afternoon, had not been introduced and had no sponsors. So for now it is a wish list, typed up the way a statute would be.
The off switch
The core of it is a rule about when the government may stop a model. Under the draft, the government could block a model’s release or use only if the model poses a foreseeable, imminent risk of more than $1 billion in damages, or of killing or seriously injuring more than 50 people. And the risk has to come from one of three places: the creation or release of a biological weapon, the model escaping its developer’s control, or a cyberattack. Anything below that line, or from some other direction, is not grounds for an order to stop.
Those numbers are not new. They come from California, where SB 53, signed last year, defines “catastrophic risk” as a foreseeable and material risk of a model contributing to the death of, or serious injury to, more than 50 people, or more than $1 billion in property damage, from a single incident. But it matters what a number is used for. In California the threshold mostly sets off disclosure: large developers have to publish frameworks for how they assess catastrophic risks, file reports and disclose critical safety incidents within 15 days. In Anthropic’s draft, the same threshold decides whether the government can do anything at all. A line that triggers paperwork is a low bar. The same line used as the minimum for an injunction is a high one.
Then there is the procedure. Independent evaluators, registered with the Commerce Department, would periodically assess each lab and its frontier models. The lab can pay them, as long as the payment doesn’t depend on what they find. (This is roughly how credit ratings work, and I am sure it will be fine.) If an evaluator finds “imminent enumerated risk,” it has 72 hours after finishing its assessment to tell Commerce. Commerce can then ask a federal judge in the District Court for the District of Columbia to suspend the model. The Daily Caller adds that a pause would need the Commerce Secretary’s written approval, that a court could halt a lab’s internal use of a model as well as its public release, and that the government could not destroy a model or its data. The draft also says outright that no other U.S. “department, agency, or officer” may suspend a model’s development, deployment or internal use.
That last clause is the interesting one, and the history explains it. Over the summer, sweeping export controls shut down two of Anthropic’s most advanced models, a decision reached in a quick series of phone calls between administration officials and company executives. Some in the tech industry worried that access to their products could be cut off whenever an official felt like it. Separately, an appeals court last month let the Pentagon blacklist Anthropic for keeping limits on Claude. Imagine, hypothetically, that your main product has been switched off by phone. When you sit down to design the next off switch, you’d want one that needs a single named agency, a written sign-off, a specific courthouse and a judge.
The states
The draft would also preempt some state AI laws, specifically those covering public disclosure of risk, third-party evaluation and reporting of critical incidents. That is awkward, because Anthropic has supported state safety laws, including two in California, and its head of U.S. state and local government relations previously told POLITICO the company hopes states will keep passing stronger and stronger laws that lift the standard nationwide. According to the Daily Caller, states would keep authority over child protection, whistleblower law and common-law liability. Preemption is a sore point on the Hill, where Democrats have rejected proposals they feared would gut rules passed in California, Illinois and New York. Industry groups say a single national standard beats a patchwork of state laws.
The rest of the draft is less controversial. Frontier labs would have to write safety frameworks, publish transparency reports for new frontier models, protect whistleblowers and report critical safety incidents to Commerce.
Not everyone is grateful for the help. Nathan Calvin, general counsel at Encode AI, wrote on X that “I found myself viscerally angry at multiple points reading this bill,” that “overall this legislation is not remotely close to meeting the moment,” and, of the intervention threshold, “This is absurd!” Anthony Cimino, Anthropic’s head of federal affairs, replied that “The threshold in our draft language pulled directly from existing standards in California law” and that it “does not absolve companies of liability.” He added: “We want a court to act quickly in an emergency. But the bar can’t be so high that the power never gets used when it’s needed.” According to the Daily Caller, he said Anthropic plans to revise the preemption language and is open to rethinking the thresholds.
Anthropic’s text is not the only one around. In the House, Reps. Lori Trahan (D-Mass.) and Jay Obernolte (R-Calif.) are leading the FRONTIER Act, which would also let the government stop risky deployments and preempt some state laws. In the Senate, Majority Leader John Thune and Sens. Amy Klobuchar (D-Minn.) and Ted Cruz (R-Texas) have been negotiating a catastrophic-risk bill. It has run into opposition from some safety groups and from Sen. Maria Cantwell (D-Wash.), the top Democrat on Cruz’s Commerce Committee, who want the government to play a bigger part in reviewing models. Rep. Greg Casar (D-Texas) and Sen. Bernie Sanders (I-Vt.) want a pause on advanced AI development and a ban on superintelligence.
The leading labs disclosed several incidents this summer in which AI agents hacked or probed websites and platforms. None reported so far appears to have caused widespread deaths or more than $1 billion in damage, so under the draft none would have counted. Last month Anthropic’s own prospectus warned that AI could pose an existential risk to humanity. That leaves a very wide range between “50 people” and “humanity,” and the bill is a proposal for who decides where in that range the government is allowed to act.