Here is the most telling line in Anthropic’s draft AI safety bill: no federal agency would be allowed to halt a model’s release on its own. Only a judge could. For a company that talks constantly about catastrophic risk, that is a notably narrow view of government power.
According to TechCrunch, citing Politico’s earlier reporting, Anthropic has been shopping the draft to Congress. A spokesperson confirmed it reflects the company’s Advanced AI Framework, which it has pushed with federal and state lawmakers since June. A release could be blocked when there is a foreseeable, imminent risk that more than 50 people could die or over a billion dollars in property damage could result. The risk must come from a biological weapon, a model escaping its developer’s controls, or a cyberattack.
Those thresholds are not new. They mirror California’s Senate Bill 53, which defines the same numbers as catastrophic risk. So Anthropic is effectively proposing to take the California template national, which would also sidestep a patchwork of state rules. That is a sensible goal for any company shipping frontier models.
The process is where it gets interesting. Independent evaluators, registered with the Commerce Department, would assess models. Labs could pay them, as long as payment doesn’t depend on the results. If evaluators find an imminent risk, they have 72 hours to notify Commerce, which must then ask a D.C. judge to stop the release. How that payment rule would hold up in practice is unclear. Financial auditing has lived with the same conflict for decades, and Demis Hassabis of DeepMind has floated a FINRA-style self-regulatory body as a different answer.
There is also a gap between the bill and what worries the industry right now. The draft focuses on mass-casualty scenarios. The more immediate fear is agentic hacking, with a run of incidents where AI programs apparently disobeyed operators and hacked sites on the open internet. A 50-death threshold does little for that kind of everyday security failure.
Still, Anthropic is ahead of its peers. Labs have mostly preferred voluntary systems, such as embedded third-party observers, over mandatory oversight. Nvidia’s Jensen Huang recently argued the industry needs no new laws at all. OpenAI and Google have favored light-touch approaches too.
Does it matter? Somewhat. A concrete draft gives lawmakers something to mark up, and that is more than most labs offer. But a regime that requires a court order, high casualty thresholds and lab-funded evaluators is one the industry can live with. Developers should read it as a signal of where the compromise might land, not as a tough new constraint.
