OpenAI says it will permit outside groups to run technical safety assessments while its AI models are being trained and evaluated, rather than limiting such testing to the period before a model’s public launch. The company describes this as an earlier, more involved form of independent evaluation during the development process.
Quartz and The Next Web both report that OpenAI is discussing arrangements with third-party organizations, including METR and Redwood Research. OpenAI has not publicly named additional partners beyond those under discussion, and it has not released specific access terms or details about how third-party testing would be operationalized. The reports also indicate OpenAI has identified priority areas for independent assessment, though the scope and timing of that work depend on how talks progress.
The outlets align on the core development: OpenAI is moving toward earlier external scrutiny of model safety during training and evaluation. They differ mainly in emphasis, with coverage focusing on the timeline shift toward in-training assessments and the fact that negotiations are ongoing and terms are not yet set.
