Anthropic has been circulating its own draft frontier AI safety bill to US lawmakers, Politico reports. The mechanism is narrow by design: the Commerce Department could ask a federal judge to halt a model, but only where the risk is imminent and the projected harm clears $1 billion in damages or 50 or more deaths or serious injuries.
The covered pathways are bioweapons, loss of control, and cyberattack. Developers would have 72 hours to report a qualifying incident.
Those thresholds are the whole argument. Set at $1 billion and 50 casualties, the bill would almost never fire — which is either the point, if you think a narrow and genuinely enforceable rule beats a broad one nobody can apply, or the problem, if you think a bar that high is a way of appearing regulated while remaining unregulated. Reasonable people land on both sides, and which side you take mostly depends on whether you read it as a floor or a ceiling.
Worth noting who is proposing it. A frontier lab writing the bill that would govern frontier labs is not disqualifying — they know the failure modes better than most legislative staff — but it is a fact readers should have when weighing the thresholds.
Sources:
Reuters via The Star ·
CryptoBriefing