Anthropic's Draft Bill Would Block AI Models Deemed Likely to Kill 50 or More People

What happened

Why it matters

• ·The 72-hour notification requirement sets a concrete disclosure clock that would, if enacted, become a hard compliance deadline for labs and their evaluators. Enterprise buyers of frontier AI services should ask vendors now whether their evaluation practices and incident notification procedures are built to meet that kind of timeline.
• ·The proposal creates a formal role for registered third-party evaluators paid by the labs they assess. This mirrors the conflict-of-interest structure that drew scrutiny in the Accenture embedded evaluator arrangement. Compliance teams conducting vendor due diligence should add evaluator independence to their assessment criteria, regardless of whether this bill passes.
• ·The bill's catastrophic-harm thresholds, including risks linked to biological weapons and uncontrolled cyberattacks, signal that regulators and legislators now treat frontier AI models as potential critical-infrastructure hazards. Organizations in regulated sectors should review whether their existing AI risk classification systems capture this category of systemic, cross-sector harm.

Governance controls affected

What to do now

• ☐Ask your AI vendors whether their current safety evaluation processes include an independent third party, and whether that party has any financial relationship with the vendor that could affect its objectivity.
• ☐Review your AI risk classification framework to confirm it includes a category for models capable of catastrophic, multi-person, or critical-infrastructure harm, not just reputational or financial risk to your organization.
• ☐Identify which frontier AI models your organization uses or plans to deploy, and map them against the proposed thresholds: foreseeable risks of mass casualties, over one billion dollars in property damage, or enabling biological or cyberattack capabilities.
• ☐Flag this proposal in your regulatory monitoring workflow so your compliance team is ready to respond quickly if the bill advances, particularly the 72-hour notification requirement, which would affect how you contract with AI vendors.
• ☐Verify that your vendor contracts require notification to your organization within a defined window if a vendor's model is flagged by an evaluator or subject to a regulatory hold, and update contract templates if this obligation is missing.

What to watch next

Compliance teams should monitor whether this draft bill attracts co-sponsors or is incorporated into broader federal AI legislation, particularly given parallel congressional activity around bipartisan AI risk frameworks. The conflict-of-interest question embedded in evaluator-paid-by-lab arrangements is likely to surface in committee hearings and could shape final bill language. If the Commerce Department receives a formal evaluator registration role, procurement teams should expect registered evaluators to become a required element of vendor due diligence for high-risk AI procurement. The EU AI Act already imposes conformity assessment requirements for high-risk systems. A parallel U.S. pre-release blocking mechanism would add a second major jurisdiction requiring pre-deployment safety verification for frontier models.