z = 6.65 sources14 articles2026-09-22 (±1 day)/en/ai/story/openai-2026-09-22
✎ Story body
How should outside bodies check whether an AI model is safe? OpenAI has put its priorities and principles for third-party assessment in writing.
What happened
OpenAI published priorities and principles for third-party assessments — evaluations of its frontier models and their safeguards carried out by external organisations. There are three pillars: the assessment should be rigorous; it should be conducted securely, so models and information do not leak along the way; and the assessor should be independent of OpenAI. Specifics such as which organisations take part and what procedures they follow are not in what we could confirm.
Why it matters
When an AI company measures its own model and declares it safe, outsiders have no way to check. Government AI safety institutes and others are experimenting with outside review, but two things pull against each other: how much of the model evaluators get to see, and how secrets are protected. Listing "secure" and "independent" side by side reads as an acknowledgment of that tension. How it works in practice is not yet settled.
What to watch
Which external bodies actually carry out assessments under these principles, and how much of their findings is made public — whether this stays a policy document or turns into published evaluation reports.