Anthropic and OpenAI have reportedly sought significant control over what outside evaluators can publish, highlighting concerns that findings could reveal trade secrets, according to reports

Friction between frontier AI developers and outside evaluators follows a familiar pattern in industries where safety assessment depends on access to the assessed: producers have historically pushed to narrow what auditors can disclose, citing intellectual property, and the resulting tension has tended to become the leverage point for formal oversight. The practical consequence in comparable episodes is that self-governed evaluation regimes lose credibility once the subjects are seen to set the publication terms, which is typically the argument that shifts scrutiny from voluntary arrangements toward mandated disclosure. The actors here matter: both firms have publicly supported third-party evaluation while seeking to protect model weights and training methods, so the distinction to draw is between legitimate trade-secret protection and editorial veto over adverse findings, since the latter is what regulators and enterprise customers have treated as disqualifying in other sectors. Watch whether major evaluators accept constrained terms, publish anyway, or walk away, and whether large customers begin demanding independent verification in procurement. For listed proxies the transmission runs through sentiment on AI-linked software names and the broader regulatory-risk premium rather than any near-term earnings channel.