An Anthropic AI safety bill is circulating on Capitol Hill before anyone has actually introduced it. According to Politico’s October 16 reporting, since corroborated by Reuters, the company has quietly shopped a draft frontier-model safety law to Congress — one that would give the federal government a narrow, court-tested power to halt model releases, but only after very large harm thresholds are met.
The draft is confidential and has not been formally introduced. That makes it a proposal, not policy — but as a signal of where one of the leading AI labs wants the regulatory line drawn, it is unusually specific.
What the Anthropic AI Safety Bill Would Allow
Under the reported text, the government could only move to block a frontier model’s release or use if its risks are both foreseeable and imminent and tied to concrete catastrophe categories: creation or release of biological weapons, cyberattacks, or AI systems escaping or evading developer controls.
The intervention mechanism and its guardrails:
- The trigger: more than $1 billion in damages, or the death or serious injury of more than 50 people.
- The referee: the Commerce Department would petition a federal judge in the U.S. District Court for the District of Columbia for an injunction.
- The scope: uniform application to all developers — competitors like OpenAI included, which Anthropic has been careful to emphasize.
Read plainly, it is a $1 billion deductible before anyone can pull the fire alarm: a deliberate attempt to write a law that captures catastrophic risk without turning every model launch into a licensing fight.
Why the Anthropic AI Safety Bill Matters Now
Context matters here. Anthropic’s public thinking on this space dates back to its October 2024 “Policy on the AI Exponential,” which proposed regulation keyed to companies spending over $1 billion on AI R&D, earning over $500 million in AI revenue, or training models above 10^25 FLOPs — a separate framework from the damages thresholds now reported, and one critics already flagged as industry-friendly line-drawing.
The draft also lands amid a crowded regulatory moment: industry standards-body talks involving OpenAI, Anthropic and Google DeepMind confirmed in September, and a broader Washington debate about whether frontier-model oversight belongs in new legislation or existing agency authority. The Anthropic AI safety bill answer is a novel one — courts, not agencies, as the circuit breaker, with a bar high enough that routine model iterations pass underneath it.
Whether Congress takes the draft as written, amends it, or shelves it, the thresholds themselves are the story. A $1 billion damages test tells you exactly how much catastrophe Anthropic considers worth federal intervention — and how much it expects the market to absorb. For coverage of the frontier labs’ wider orbit, see our reads on Claude’s NSA seed-mystery work and Anthropic’s Jack Clark on AI’s $44 trillion GDP scenario.
