Anthropic Promises to Revise Draft AI Safety Bill After Backlash Over 37-Page AEGIS Text

What happened

On Oct 16, Politico reporter Owen Dahlkamp reported that Anthropic had privately circulated a 37-page discussion draft of a federal AI safety bill, working title AEGIS Act, to some members of Congress. Anthropic did not deny the report and confirmed the draft comes from the Advanced AI Framework it has promoted since June.

Opposition erupted on Oct 19–20 after the leak. Encode's general counsel Nathan Calvin said reading the draft made him 'shake with anger' and that passing it would leave less protection than the status quo; NY assemblymember Alex Bores (author of the RAISE Act) and Transformer author Shakeel Hashim also spoke out publicly.

OpenAI's chief global affairs officer Chris Lehane responded publicly on Oct 19: 'No single company should write the rules for such an important technology.' Former White House AI lead David Sacks criticized from the opposite direction, saying the draft would sweep in most Silicon Valley companies.

Key facts

1. Pause threshold: Only when a model poses a 'foreseeable and imminent risk' — more than 50 deaths or serious injuries, or more than $1 billion in losses — and the risk stems from bioweapons, loss of model control or cyberattacks, could the Commerce Department ask a D.C. federal judge to order a model paused; the figures mirror California's SB 53.
2. Reporting mechanism: The draft would require evaluators to register with the Commerce Department and report risks within 72 hours.
3. State-law clause: The draft would also preempt some state AI rules, including risk-disclosure, third-party evaluation and incident-reporting requirements — the most contested point of the backlash.
4. Anthropic's response: Cimino said on X the draft was shared with only 'a handful of congressional offices'; acknowledged the preemption language was 'imprecise' and promised a rewrite clarifying states' roles in evaluator standards and lab-shutdown authority; admitted the 50-person/$1B threshold came from California law and that 'now is the time to re-evaluate the threshold'; new proposals on evaluator standards and stricter incident reporting are also coming.