OpenAI published principles for external assessments of its AI systems on September 22, 2026. They cover safety cases, safeguards, capabilities and investigations of serious model misalignment.
The post calls for agreed assessment scopes, proportionate access and conflict-of-interest rules. Assessors should be able to stand behind their own findings while respecting agreed confidentiality and security limits.
The AI Evaluator Forum had proposed its own minimum conditions on September 18. Updated on September 23, its letter calls for transparency about access and funding, protection against retaliation and direct communication with company boards. Signatories act personally, not automatically on behalf of their employers.
OpenAI's framework is not tied to individual launches. Assessments may take weeks or months; the post names no selected partners. The access and publication rights actually negotiated will determine how much scrutiny it enables.