Calm down. · Both
Calm down: Anthropic didn't set a body count. It wants to move the off-switch.
AI Weekly's headline: "Anthropic Floats AI Safety Bill With $1B, 50-Death Thresholds." Calm down.
Per Reuters, relaying Politico, Anthropic quietly pitched lawmakers a draft bill. The government could block a model only if it poses "a foreseeable, imminent risk of more than $1 billion in damages, or of killing or seriously injuring more than 50 people." Those numbers aren't Anthropic's. The House FRONTIER Act (H.R. 9925) uses the same thresholds: "the death of, or serious injury to, more than 50 people, or more than $1,000,000,000 in damage to, or loss of, property." Anthropic's own framework points to California's SB 53 for the same idea. And this is a confidential draft. No bill number, no named sponsor, and every detail comes from Politico's reading of a draft that hasn't been published.
The real change is who holds the plug. Under the FRONTIER Act, the Commerce Secretary "may issue an emergency order" halting a model. Under the draft, as reported, Commerce would have to ask a federal judge in D.C. to halt a model, and every other agency is explicitly barred from suspending one. Evaluators register with Commerce. The labs can pay them, as long as the payment isn't tied to the outcome. Those evaluators must report "imminent enumerated risk" to Commerce within 72 hours.
Anthropic's spokesperson says the bill "represents the proposals outlined in Anthropic's Advanced AI Framework." Partly. The framework says the government should be able to block models that pose "a significant risk of catastrophic harm"; the draft, as reported, says imminent. The framework lists four risk categories; the draft's halt trigger, as reported, covers three and leaves out automated R&D. The framework also warns about "evaluator shopping," where companies pick whichever evaluator "will ask the least of them." The draft lets labs pay their evaluators.
My take: this is the regulatory capture I already complained about. A lab is helping write the rules it gets judged by, and choosing who gets to pull the plug. Lab-paid evaluators are the legal tripwire. We just watched a paid tester leave the door open.
What to do: Nothing yet. Anthropic's framework would cover only developers training above 10²⁵ FLOP with more than $500 million in AI revenue or $1 billion a year in R&D, which excludes almost everyone reading this. If you buy AI evaluations yourself, write independence and egress into the contract.
Sources
- Reuters via KELO: Anthropic quietly pitches its own AI safety bill to lawmakers, Politico reports (Oct 16, 2026)
- Bloomberg via Moneycontrol: Anthropic pitches AI safety bill to lawmakers (Oct 16, 2026)
- Anthropic's Advanced AI Framework (June 2026)
- H.R. 9925, FRONTIER Act (Jul 23, 2026)
- AI Weekly (Oct 17, 2026)