OPENAI AND ANTHROPIC WANT TO GRADE THEIR OWN HOMEWORK

Reuters reported that OpenAI and Anthropic, while publicly calling for independent reviews of frontier models, privately pushed for significant control over what outside safety evaluators could publish. Two people familiar with the talks said that as recently as October 5 both companies sought strict editorial authority, citing risks that findings could reveal trade secrets.

Anthropic disputed that seeking significant control characterised all its discussions, saying established evaluators control their methods, criteria, conclusions and publication, while it reserves comment and approval rights for new evaluator firms. OpenAI said it was not seeking approval rights over the full text of reports, that its standard practice is to redact confidential information, and that evaluators can disclose whether major redactions were made. Experts warned that lab control over access and publication could undermine evaluator independence.