Anthropic shops draft AI safety bill tying model blocks to 50-death risk and judge approval
Anthropic is circulating a draft congressional AI safety bill that, according to Politico reporting cited by TechCrunch, would restrict model distribution when independent evaluators flag foreseeable imminent catastrophic risk. Covered scenarios include biological weapons, models evading developer controls, or cyberattacks where more than fifty deaths or one billion dollars in property damage could result, thresholds similar to California Senate Bill 53. Reported mechanics would route halts through the Commerce Department and a federal judge in Washington, D.C., while barring federal agencies from blocking release on their own. Commerce-registered third parties would evaluate labs periodically. Critics call the leak weaker than the FRONTIER Act, and the packet describes only a pitched draft, not enacted law, leaving enforcement details uncertain.
Anthropic shops draft AI safety bill tying model blocks to 50-death risk and judge approval
Politico reported Anthropic is circulating a congressional draft that could restrict model distribution when independent evaluators flag foreseeable imminent catastrophic risk, including cases where over 50 people could die or over a billion dollars in property damage could result. The draft would route halts through the Commerce Department and a D.C. judge and would bar federal agencies from blocking release on their own, according to the report.
Key takeaway
Anthropic's leaked draft sketches Commerce-and-judge gating at high catastrophic-risk thresholds, but the packet frames it as lobbying, not federal law in force.
What happened
Politico reported that Anthropic has been shopping a draft AI safety bill to Congress that could block model distribution when evaluators see a foreseeable, imminent catastrophic risk, including more than fifty deaths or over one billion dollars in property damage, per TechCrunch's account of the reporting.
Reporting summarized in the packet says halts would go through the Commerce Department and a D.C. federal judge, federal agencies could not block release alone, and Commerce-registered independent evaluators would periodically review labs; Anthony Cimino said the text translates Anthropic's Advanced AI Framework and was shared with a small number of congressional offices.
Evidence
Politico reported Anthropic is circulating a congressional draft with catastrophic-risk thresholds and a Commerce-and-judge halt process.
TechCrunch AI · attributed
Politico reported Anthropic is circulating a congressional draft that could restrict model distribution when independent evaluators flag foreseeable imminent catastrophic risk, including cases where over 50 people could die or over a billion dollars in property damage could result.
The draft would bar federal agencies from blocking model release on their own without the described judicial route.
TechCrunch AI · attributed
The draft would route halts through the Commerce Department and a D.C. judge and would bar federal agencies from blocking release on their own, according to the report.
Politico-linked reporting describes Commerce-registered third-party evaluators and high harm thresholds for blocking release.
Hacker News AI · attributed
According to Politico, an independent third party registered with the Department of Commerce would periodically evaluate the lab and its AI models.
AI safety advocates criticized the leaked draft as weaker than the bipartisan FRONTIER Act.
Techmeme · attributed
Anthropic's leaked federal AI bill draws widespread criticism from AI safety advocates for being significantly weaker than the bipartisan FRONTIER Act
Anthropic federal affairs lead Anthony Cimino said the draft translates the Advanced AI Framework and was shared with select congressional offices.
LessWrong · attributed
This draft is a translation of those ideas into legislative language, shared with a small number of congressional offices that have been working on these issues.
Why it matters
If similar federal language advances, frontier model release planning could hinge on periodic third-party evaluations and judicial blocking orders tied to explicit catastrophic-risk triggers.
Limits and uncertainties
The packet consistently describes a draft being pitched or circulated, not legislation enacted or introduced with confirmed congressional uptake.
Excerpts are incomplete and do not establish full bill text, evidence standards for halts, or how the fifty-death and one-billion-dollar thresholds would be measured in practice.
Practical implications
Operators should treat Commerce-registered evaluator requirements and seventy-two-hour reporting mechanics described in Politico-linked excerpts as speculative until formal legislative text and agency rules exist.
Builders tracking state and federal preemption should note LessWrong commentary that draft preemption language may not preempt state model suspensions as written, an interpretation the packet does not verify as authoritative.
What to watch
Whether any congressional office introduces federal AI safety text matching the leaked Anthropic draft or the bipartisan FRONTIER Act framework cited by critics.
Further published bill text clarifying who initiates blocking orders, what triggers Commerce action, and whether federal agency halt authority remains constrained as TechCrunch summarized from Politico.
