The AI Industry

Anthropic’s Draft AEGIS Act: What the 50-Person Threshold Would and Would Not Let Washington Stop

A leaked 37-page draft would let a court halt an AI model only above $1 billion in harm or 50 deaths. Here is how it works, who objects and what is unsettled.

Elena Vasquez

Former ML Researcher, Industry Analysis Lead

Published 7 min read
Stunning view of the US Capitol in Washington DC during a beautiful sunset.
In this story 10 sections

A leaked draft called the AEGIS Act, tied to Anthropic's policy framework, would let a federal court halt an AI model only for an imminent risk of over $1 billion in damage or more than 50 deaths or serious injuries. It has no sponsor and is not law. Critics say the bar is too high.

This explainer is for readers who want the actual mechanics instead of the headline. It covers what the draft would do, who would decide, what it would override in state law and what its critics and defenders say. Emergent Wire covers the AI industry, and this is a good example of a story where the details change the meaning of the headline.

What Does the Leaked Draft Say?

Interior view of an American courthouse in Kirksville, Missouri, featuring a judge's desk and flags.

The draft is a 37-page bill dated September 29 with no sponsor. Politico published it as an exclusive on October 16, and Anthropic confirmed it reflects the company's Advanced AI Framework, according to Let's Data Science, which summarized the text. It is a discussion draft, not an introduced bill.

The government could halt a model only if it poses a foreseeable and imminent risk of more than $1 billion in damage, or of killing or seriously injuring more than 50 people. The risk must come from a biological weapon, a model evading its developer's controls, an AI-enabled cyberattack, or AI research that substantially speeds those up.

The draft also sets routine duties. Covered developers would need safety frameworks and would have to report incidents within 15 days, or within 24 hours if death or serious injury is imminent.

Who Would Decide Whether to Halt a Model?

A chain of actors would. Independent evaluators registered with the Commerce Department would assess covered developers, and an evaluator that finds a present or impending listed risk would have 72 hours to notify Commerce. Labs may pay evaluators, as long as payment does not depend on the results.

Only the Commerce Secretary could seek a blocking order. The Secretary must issue a written determination and petition the U.S. District Court for the District of Columbia, which would have exclusive jurisdiction, according to the same summary.

A court order would last a bracketed 45 days and could be renewed once. It could not force a lab to delete weights, data or code. The judge would decide without deference to Commerce or the evaluators.

Why Is the 50-Person Threshold So Controversial?

A professional individual signs legal documents at a desk in an office setting.

Critics say the number sets the bar for action very high. Under the draft's logic, a model that could cause 49 deaths or $999 million in damage would not meet the standard. That reading is the core of the criticism, and it is a fair reading of how a numerical threshold works.

Anthropic's Anthony Cimino, who leads federal affairs, defended the figures as drawn from California law and said they do not absolve companies of liability. He also said it is probably time to re-evaluate thresholds for a national standard.

There is a second concern. Critics flagged that the court must weigh the harm a block would cause the developer, users and the public, with no matching instruction about the harm of letting a dangerous model keep running. That tilt, they argue, makes a halt even less likely.

The table below sets out the draft's main triggers and where objections arise.

Why Is the 50-Person Threshold So Controversial?
FeatureWhat the draft saysMain objection
Harm thresholdOver $1 billion or 50 peopleBar is too high
EvaluatorsRegistered with Commerce, paid by labsRisk of self-certification
Court roleD.C. court, 45-day ordersWeighs developer harm heavily
ScopeLarge developers at 10^25 operationsToo broad, say some
State lawsSection 14 preempts severalWeakens state rules

What Would the Draft Do to State AI Laws?

Section 14 would preempt state laws on developer safety disclosures, third-party evaluation and incident reporting. That would reach California's SB 53, New York's RAISE Act and Illinois' SB 315. Child safety, consumer protection and civil rights laws are carved out.

The preemption fight echoes another draft we covered, the Great American AI Act draft and its three-year state preemption. In both cases, the question is whether a single national rule should replace a growing patchwork. Our overview of state AI laws in 2026 shows how large that patchwork has become.

Cimino said the preemption language was not precise enough and would be revised to be narrow. Emergent Wire's editors would watch whether the revised text keeps the same list of covered state laws, since the list decides the real effect.

How Does the Compute Threshold Compare?

Black and white photo of a traditional legislative assembly chamber, featuring empty seats and desks.

The draft defines large developers at 10^25 training operations. That is ten times below California's SB 53 line, so more companies would fall under the federal rules. Training operations are a rough measure of how much computing power went into building a model.

The lower bar cuts both ways. It brings more developers into the safety framework, which supporters of regulation like. But David Sacks argued that the lower threshold would sweep in much of Silicon Valley, which is a complaint that the bill goes too far.

That split shows how unusual this fight is. Some critics say the bill is too weak on harm, while others say it is too broad on scope. Both can be true of different provisions.

Who Objects, and What Do They Want Instead?

Several safety advocates and lawmakers have pushed back. Nathan Calvin, general counsel at Encode, said the bill is not remotely close to meeting the moment and argued it makes it harder than existing law for a court to block a frontier model.

California Senator Scott Wiener, who wrote SB 53, said the draft falls short on public safety, sets no minimum standards for safety plans or auditors and weakens emergency shutdown powers. Charlie Bullock of LawAI warned that without licensing, auditors may compete by offering the least burdensome audits. Shakeel Hashim of Transformer said evaluators could effectively self-certify.

Anthropic says it is weighing how to prevent evaluator shopping and plans to publish proposals on evaluator requirements. Whether that satisfies critics is unknown.

How Does It Compare With Other Federal Proposals?

Researchers in lab coats analyzing experimental notes during a scientific study.

It is not the only bill. The bipartisan FRONTIER Act, from Representatives Obernolte and Trahan, was introduced in July 2026 and would also let the government stop risky deployments. In the Senate, Majority Leader Thune is negotiating a catastrophic-risk bill with Senators Klobuchar and Cruz, according to the summary we reviewed.

A formal bill would also face a crowded field of federal scrutiny. We covered one example in our report on FTC civil investigative demands to OpenAI and Anthropic. Regulators and lawmakers are looking at the same companies from several directions at once.

Lawmakers' responses to the draft are unclear, according to Let's Data Science coverage of the pitch. A discussion draft circulated to lawmakers is a long way from a law, and many drafts change or disappear before introduction.

Why Did Anthropic Back a Bill Critics Call Weak?

The draft turns Anthropic's public framework into statutory language, and the company calls those policies sensible. Its framework covers biological weapons, offensive cyber operations, loss of control and automated AI research, but the circulated bill narrows when the government can step in directly.

Supporters of the approach argue that a narrow, hard-to-trigger power is easier to pass and harder to challenge in court. A bill that survives Congress and the courts may do more than a stronger one that never becomes law. That is a real trade-off, not a dodge.

Critics reply that the design lets labs shape the rules they will follow, since the same companies fund the evaluators and wrote the template. The Washington Post, as summarized in coverage we reviewed, noted the draft differs from the FAA-style model Anthropic's chief executive advocated this summer. Both views will get tested if a sponsor steps forward.

What Should Readers Watch Next?

Watch for three things. First, whether Anthropic publishes revised text, especially on preemption. Second, whether a sponsor introduces the bill in either chamber. Third, whether evaluator standards get written into any version, since critics focus on that gap.

Keep the limits of the reporting in mind. The draft is confidential and the terms here come from press summaries of a leaked text, so details may differ from the final language. Readers should check the bill text itself if it is introduced.

A fair counterpoint is that a narrow trigger can still add value. Routine evaluations, incident reporting and safety frameworks create information the government does not have today. Supporters argue that is a useful floor even if the emergency halt power is rarely used.

Conclusion

The AEGIS Act draft shows how a few numbers can define an entire regulatory approach. A halt would require over $1 billion in harm or more than 50 people hurt or killed, and a court, not an agency, would make the call. State preemption and evaluator independence are the next fights. Follow whether a sponsor emerges and whether the revised text changes those two pieces.

What is the AEGIS Act draft?
It is a leaked 37-page discussion draft, dated September 29, that Anthropic confirmed reflects its Advanced AI Framework. It has no sponsor and is not law. Politico published it on October 16, 2026. It would let a federal court halt a model only for imminent risks above set harm thresholds.
What is the 50-person threshold in the Anthropic draft bill?
Under the draft, the government could halt a model only if it poses a foreseeable, imminent risk of more than $1 billion in damage or of killing or seriously injuring more than 50 people. Critics say the bar is too high. Anthropic says the figures come from California law and that thresholds may be re-evaluated.
Would the draft override state AI laws like California SB 53?
Section 14 would preempt state laws on developer safety disclosures, third-party evaluation and incident reporting, reaching California SB 53, New York RAISE Act and Illinois SB 315. Child safety, consumer protection and civil rights laws are carved out. Anthropic says the language will be revised to be narrower.
Who can ask a court to halt an AI model under the draft?
Only the Commerce Secretary could do so, after issuing a written determination and petitioning the U.S. District Court for the District of Columbia. Independent evaluators registered with Commerce would flag risks within 72 hours. A court order would last a bracketed 45 days and could be renewed once.

Written by

Elena Vasquez

Former ML Researcher, Industry Analysis Lead

9+ years experience

Elena Vasquez holds a PhD in computer science with a research focus on model evaluation methodology, and spent seven years split between an academic lab and a mid-size AI startup's research team before moving into journalism full time. She joined Emergent Wire to bring a practitioner's skepticism to industry coverage — she has built and broken enough benchmarks herself to recognize when a vendor's claimed number is doing more marketing work than measurement work. Her beat covers model capability claims, evaluation methodology, and the gap between published benchmarks and production performance.

Credentials

  • PhD, Computer Science (Machine Learning)
  • Former research scientist, applied AI startup (4 years)
  • Published research on model evaluation methodology