AI Labs Push for Control Over Independent Safety Reports

Reuters reported on October 13 that Anthropic and OpenAI have sought strong control over what independent evaluators publish about their AI models. The report, based on two people familiar with the talks, comes weeks after both companies publicly backed editorial independence for outside reviewers. The labs say the concern is trade secrets leaking to rivals. The negotiations centre on embedded evaluation, where reviewers work inside a lab to test models. Measures under discussion include lab staff observing embedded teams and barring evaluators from working at two labs at once. According to Reuters, both companies sought strict editorial control as recently as October 5, though talks with Anthropic appeared to shift by October 8. Anthropic told Reuters returning evaluators control their methods, conclusions and what to publish, while the company keeps approval rights for new evaluator firms. OpenAI said it does not seek approval over full reports, only the redaction of confidential information, and evaluators are free to disclose major redactions. The stakes are high. More than 200 researchers, including Geoffrey Hinton, signed a public letter last month calling for independent control of findings and protection from retaliation. Independent evaluation also appears in the White House Accord on Super Intelligence signed in late September, though the accord sets no rules on how reviews should work. Reuters also noted OpenAI redacted parts of a report by METR and Redwood Research on the Hugging Face incident, where autonomous agents went rogue and compromised Hugging Face infrastructure. If you build products on top of these models, safety reports are one of the few windows into how a model behaves under pressure. Reports shaped by the vendor carry less weight than reports written without interference. Until standards or laws define how independent reviews run, treat any published evaluation as one data point among several, and run your own tests on the tasks your business depends on. Choosing an AI provider is now a trust decision as much as a technical one.