News · SI
Anthropic has been quietly showing lawmakers a draft federal bill on frontier AI safety, POLITICO reports. The government could stop a model only for a foreseeable, imminent risk of more than 50 deaths or serious injuries or more than $1 billion in damage, and only the Commerce Department, through a federal judge, could do it.
Anthropic representatives have been quietly showing lawmakers a draft federal bill on frontier AI safety and security, POLITICO reported on Friday US time, citing three people with knowledge of the effort[1]. The bill would let the government block a model's release or use only if it posed a foreseeable, imminent risk of more than $1 billion in damage, or of killing or seriously injuring more than 50 people[2].
An Anthropic spokesperson told Reuters the bill represents the proposals in the company's Advanced AI Framework, published earlier this year[2]. "These are sensible policies we believe Congress should act on," the spokesperson told Bloomberg[3].
01 / 05
The damage threshold isn't enough on its own. As reported, the risk also has to come from one of three sources: a model creating or releasing a biological weapon, a model evading its developer's controls, or a cyberattack[2].
That's narrower than Anthropic's published framework, which names four risk categories for developer obligations: biological weapons, offensive cyber operations, loss of control, and automated research and development that could accelerate the other three[4]. Automated R&D doesn't appear in the reported halt trigger.
02 / 05
Enforcement would run through the Commerce Department. Independent evaluators would register with Commerce and periodically assess each lab and its frontier models[2]. Labs could pay those evaluators, as long as payment isn't tied to the result[2].
If an evaluator finds an "imminent enumerated risk", it would have to tell Commerce within 72 hours of finishing its assessment. Commerce could then ask a federal judge in the District of Columbia to suspend the model, and the draft explicitly bars any other US department, agency or officer from suspending a model's development, deployment or internal use[2].
That mirrors the framework's "safeguards against overreach", where Anthropic suggests an agency should only pursue remedies by going to court rather than imposing them directly, and should act only on specific listed violations, not its own general assessment of risk[4].
03 / 05
The reports don't spell out which companies the bill would reach. The framework it's based on targets developers that train models using more than 10^25 FLOP of compute and either earn more than $500 million a year in AI revenue or spend more than $1 billion a year on AI research and development[4]. It also calls for publishing catastrophic-risk test results, reporting critical safety incidents within 15 days, whistleblower protections and civil penalties that scale with revenue[4].
On state law, the framework says Congress shouldn't preempt states unless it passes a federal regime at least as strong as Anthropic's proposals, and even then only for the specific functions it takes over[4]. The Reuters and Bloomberg accounts don't say how the draft bill handles that question.
05 / 05
This is a frontier lab writing the rules it would like to be held to, and the shape is clear: real testing and outside evaluators, but a very high and very specific bar before anyone can stop a model, with one agency and one court holding that power. For teams building on Claude, GPT or Gemini, nothing changes yet. The thing to watch is whether a sponsor picks it up, and whether the 50-person and $1 billion thresholds survive contact with Congress.
Sources
FAQ
It's draft legislation on frontier AI safety and security that Anthropic representatives have been showing to lawmakers on Capitol Hill, according to POLITICO, which reported it on 16 October 2026 (US time) citing three people with knowledge of the effort. The text hasn't been published, and the reports don't name a sponsor.
Only when a model poses a foreseeable, imminent risk of killing or seriously injuring more than 50 people or causing more than $1 billion in damage, according to the reports. The risk also has to come from one of three sources: creating or releasing a biological weapon, the model evading its developer's controls, or a cyberattack.
The Commerce Department. Independent evaluators would register with Commerce and periodically assess labs and their frontier models, and would have to report an imminent risk within 72 hours of finishing an assessment. Commerce could then ask a federal judge in Washington, DC to suspend the model. The draft reportedly bars every other federal agency or official from halting a model's development, deployment or internal use.
Yes, as long as the payment isn't tied to the outcome of the evaluation, per the reports. Anthropic's published framework goes further on this point, warning about evaluator shopping and suggesting the government could rate evaluators or fund them directly.
A spokesperson told Reuters and Bloomberg the bill represents proposals in Anthropic's Advanced AI Framework, which it has been pushing with federal and state lawmakers since June. The spokesperson said it would require developers to test their models, be transparent about findings and get independent evaluations, and would give the government the ability to block or deter dangerous deployments of the most powerful models.
Because it sets a high bar for government action: as reported, a halt requires a risk of more than 50 deaths or serious injuries or more than $1 billion in damage, so a risk exceeding neither threshold wouldn't qualify. Anthropic's own framework frames tight limits like these as a guard against regulators overreaching, arguing for court enforcement and narrow agency discretion.
That’s the story
Spectrum Web Co ships SI and Web3 products, places talent and builds social narratives as monthly services.