Anthropic have released a targeted transparency framework proposal applying only to the largest AI system developers, with suggested thresholds including annual revenue of $100 million or R&D expenditures of $1 billion annually.
The framework establishes six core requirements for covered frontier model developers. Companies must implement a Secure Development Framework assessing risks including chemical, biological, radiological and nuclear harms, plus misaligned model autonomy risks. This framework must be publicly disclosed on company websites with self-certification of compliance, subject to reasonable redactions for sensitive information.
Additional requirements include publishing system cards summarising testing procedures, evaluation results and mitigations at deployment, with updates for substantial model revisions. The proposal explicitly prohibits false statements about framework compliance, creating legal violations that enable existing whistleblower protections. Anthropic notes this approach mirrors existing practices from leading labs including Google DeepMind, OpenAI and Microsoft, which have implemented similar approaches.
The company emphasises the framework's lightweight design to avoid impeding AI innovation or slowing benefits like "lifesaving drug discovery, swift delivery of public benefits, and critical national security functions." Anthropic states that rigid government-imposed standards would be counterproductive given evaluation methods become outdated within months due to technological change pace.
The proposal could be applied at federal, state or international levels. Anthropic acknowledges that startup community input is needed for appropriate threshold determinations, with periodic reviews as technology and industry landscapes evolve. The framework aims to provide interim safety measures while industry, governments and academia develop comprehensive evaluation methods and agreed-upon safety standards.
The transparency requirements target frontier model developers building the most capable systems, establishing baseline accountability standards while enabling public and policymaker distinction between responsible and irresponsible development practices. Smaller developers receive appropriate exemptions to avoid burdening the startup ecosystem.
The framework standardises existing industry best practices without setting them permanently, ensuring voluntary disclosures cannot be withdrawn as models become more powerful. Public visibility into safety practices preserves private sector agility while providing policymakers evidence to determine if further regulation is warranted.