OpenAI and Anthropic face scrutiny over AI safety reports

Independent evaluation only works if outsiders can report uncomfortable findings. The strategic issue here is who controls the evidence used to judge frontier systems, rather than whether a particular model passes a test. Reuters describes disputed negotiations over publication rights and confidentiality, making the final contractual terms more important than public assurances of independence. ([cnnbrasil.com.br](https://www.cnnbrasil.com.br/economia/money/inteligencia-artificial/anthropic-e-openai-buscam-controlar-relatorios-de-seguranca-de-ia/))

For Race to AGI readers, my assessment is that evaluation access should be treated as infrastructure: it determines how reliably capability gains and dangerous behavior become visible. Confidentiality can protect legitimate research secrets, but an agreement that lets a laboratory suppress material findings would weaken the value of outside scrutiny. Competitively, credible publication rights could become a differentiator for laboratories seeking enterprise customers and government trust. Conversely, restrictive arrangements could make comparisons between laboratories less meaningful. This report does not establish that AGI development has accelerated or slowed. It raises a practical question about whether the institutions measuring progress can remain independent of the companies setting its pace.