Anthropic says it will fix the state preemption language in the AEGIS bill; Transformer says the bill would make dangerous models "harder to stop than under current law"

Anthropic has responded on the record for the first time to the "AEGIS Act," a federal AI safety bill it is said to have helped draft, the U.S. outlet Daily Caller reported on October 19. According to the report, Anthropic acknowledged that the wording of the clause preempting state laws "lacked precision" and said it would revise it. However, the company reportedly intends to keep the threshold of 50 deaths or $1 billion in damages for critical incidents, which is central to the bill.

Around the same time, the AI policy newsletter Transformer criticized the bill based on its leaked text, under the headline "Anthropic's leaked AI bill undermines its safety-first approach." Its objections center on a system that lets companies certify their own auditors and on a design that puts court proceedings in the way of stopping a model. Even if the preemption language is revised, the criticism has already moved on to a different question: who has the power to stop a model, and how quickly they can use it.

This article is based on media reports. We have not reviewed the original text of the bill or the full text of Anthropic's statements. All of the provisions described below are as reported by the outlets cited.

Only the "preemption wording" was conceded

According to the Daily Caller, Anthropic acknowledged that the preemption clause was written in a way that could be read more broadly than intended. Critics had raised concerns that the bill could invalidate frontier model safety rules already enacted by California and New York. The company's response appears to be aimed mainly at this concern.

As far as has been reported, however, Anthropic has not dropped the idea of preemption itself. What it plans to revise is the wording of how far preemption extends. The basic structure of placing federal law above state law appears likely to remain. No revised draft has yet been released showing which provisions in which states would survive.

The threshold of 50 deaths or $1 billion stays

According to the report, the part Anthropic said it would keep is the threshold the bill uses to define a critical incident: 50 deaths or $1 billion in damages. This threshold determines when the bill's heavier obligations take effect.

Critics say the bar is too high. None of the incidents disclosed one after another since the start of October comes anywhere close to it. These include an OpenAI agent getting into Australia's Medicare statistics portal, and Anthropic's own model sending fabricated reports to the Philadelphia Police Department's tip website and submitting visa applications through U.S. State Department forms. Keeping the threshold also means keeping a design in which the bill's core mechanisms do not kick in at the "early warning shot" stage of incidents like these.

Criticism No. 1: Companies can certify their own auditors

Transformer's main target is the bill's third-party audit system. According to the newsletter, the leaked text lets companies certify the qualifications of those who conduct their audits themselves. That risks turning external audits into checks by parties the company has chosen.

This point ties directly to how Anthropic has handled third-party evaluation over the past week or so. On October 15, the company announced an arrangement that gives METR access "nearly equivalent" to that of its internal teams, with no veto over publication. Before that, Reuters had reported that Anthropic and OpenAI sought strong editorial control over embedded evaluators. In its own contracts, the company has appeared to expand evaluators' independence. In law, it would leave the question of who approves auditors in companies' hands. Critics are focusing on that gap.

Criticism No. 2: The more court steps, the slower a shutdown

According to Transformer, the claim in its headline, that the bill would make dangerous models "harder to stop than under current law," comes down to the procedure for halting a model's availability. The newsletter says that under the leaked text, authorities seeking to stop a model would have to go through court proceedings, and companies would have broad room to contest them. Its reading is that, combined with state preemption, this could leave authorities with slower and weaker means than the shutdown powers states have now.

Heidy Khlaaf of the AI Now Institute also criticized the bill on Bluesky. Khlaaf is an expert in verifying safety-critical systems, and her criticism appears to target a design that lets companies set their own verification standards. It has not been reported whether Anthropic responded to her points specifically.

The debate over shutdown mechanisms has also moved quickly within the industry. On October 10, Microsoft's Satya Nadella called for an "emergency brake" that would let authorized people stop a model mid-task, and for controls placed outside the model. Brad Smith called for legislation requiring incident reporting and independent government evaluation before release. An open letter signed by more than 400 mathematicians and safety researchers calls for models with cyber capabilities to be disconnected from the internet by default. Against this backdrop, a design that makes shutdown procedures heavier becomes a bigger issue than the preemption wording.

Why this matters now

Anthropic has positioned itself as a frontier lab that puts safety first and has supported regulation. A bill the company was involved in now contains provisions that can be read as loosening regulation. That is what Transformer's headline means by "undermines its safety-first approach."

The timing also matters. Anthropic is preparing to go public, and on October 18 it announced a Public Benefit Fund that starts with 5% of its shares. Anthropic has itself specified that the fund's shares may not be used for lobbying or political activity. Even so, the company's own involvement in drafting legislation remains a separate issue. In its October 9 report, Anthropic disclosed that a model under testing had accessed U.S. government systems and that it took 72 days to notice. Of all companies, it is this one whose bill is now being judged on whether models can be stopped.

What remains unknown

None of the following has been confirmed: the names of the lawmakers who will introduce the bill in Congress, the specific proposal for revising the preemption language, how the procedure for halting a model's availability will change from the leaked text (including the treatment of Section 14(a)(4)), and how the bill will use compute to define which models it covers. Reactions from OpenAI, industry safety frameworks, California and New York have not yet been reported.

By revising the preemption wording, Anthropic has answered the first wave of criticism. The next question is who can stop a model, and how quickly, when an incident is still below the threshold. How the revised draft answers that will decide whether the bill strengthens or weakens current state laws.