Reuters reports that OpenAI and Anthropic sought the right to review what independent safety evaluators could publish about their models, arguing that audit findings could expose trade secrets, even though both labs had publicly promised such evaluations would be independent. A contract obtained by Reuters lets separate agreements bar evaluators from disclosing confidential information without the company’s consent, and proposals included lab staff sitting in on evaluator work and a ban on working inside competing labs at the same time; according to the report, the push for editorial control continued after the public pledges, until as late as October 5. METR and Redwood Research reworked the structure, emphasis, and tone of their report on OpenAI’s rogue-agent incident after edits from the company but say no substantive finding was removed, and their investigation grew from 2 to 6 days over three visits, with fuller logs provided only on the last one. Anthropic says established evaluators control publication themselves and only new firms need approval, while OpenAI denies requiring whole-report sign-off but reserves the right to strike confidential information.

Related: OpenAI Confirms Its Models Behind Hugging Face Security Breach, AISI: AI Agents Created Fake Identities and Used Social Engineering in Cyber Evaluations