Anthropic and OpenAI sought strict control over findings published by third-party AI safety evaluators due to fears of trade secret leaks to commercial rivals. Despite public pledges supporting independent review, contract negotiations placed labs in direct friction with evaluators demanding unvetted publishing rights. While the White House Accord on Super Intelligence mandated third-party evaluations, it lacked binding execution standards.
AI companies' stances on publication control
- ▪Contractual language used by major AI companies bars external evaluation firms from disclosing confidential information in public reports without prior consent
- ▪Anthropic CEO Dario Amodei and OpenAI publicly stated in late September 2026 that external safety evaluators should maintain independent right to publish findings without company editorial control
- ▪As recently as October 5, 2026, OpenAI and Anthropic sought strict editorial control over what third-party safety evaluators could publicly report about their AI models
- ▪OpenAI stated that its standard practice is to redact confidential information rather than seek full text approval, allowing evaluators to publicly report whether major redactions occurred
- ▪Anthropic stated that it allows existing evaluator partners to control their methods, criteria, conclusions, and publication content, but reserves comment and approval rights for new evaluator firms such as Faculty
Evaluator trade secret concerns and measures
- ▪Measures considered by AI developers to guard trade secrets include assigning employees to observe embedded evaluation teams and prohibiting evaluators from working simultaneously at competing laboratories
- ▪OpenAI and Anthropic raised concerns that embedded evaluation staff working on-site could inadvertently or intentionally leak proprietary intellectual property to commercial rivals
Anthropic's embedded evaluator partnership
- ▪Anthropic designated Accenture unit Faculty as its first embedded evaluator in late September 2026, with Faculty CEO Marc Warner declaring that evaluation transparency and independence are critical for public trust
- ▪Anthropic stated in late September 2026 that specific details regarding what information embedded evaluators can access and how they report findings were still being worked out
White House Accord on Super Intelligence
- ▪The White House Accord on Super Intelligence omitted procedural standards or compliance mechanisms dictating how independent AI safety reviews must be conducted
- ▪The White House Accord on Super Intelligence signed by President Donald Trump and AI executives in late September 2026 mandated independent evaluation as one of four core safety measures
Debatable claims
- ▪Observing embedded safety evaluators goes too far in restricting independent oversight
- ▪AI developers should retain right of refusal over third-party evaluator publications
- ▪Corporate redactions undermine the validity of independent AI safety reports
Story comments
Loading comments…