OpenAI says it will open frontier models to outside safety assessors
The company said it is setting out four priority areas, and that assessors should be able to challenge its assumptions and reach their own conclusions.
OpenAI said on Tuesday it will support independent assessments of its models, with access across training, evaluation and deployment. The company said that access should let third-party assessors challenge its assumptions and identify risks it may have missed.
Bloomberg reported that OpenAI plans to let third-party groups conduct technical safety evaluations during the training, evaluation and deployment phases, according to a Techmeme summary of the story. OpenAI published a page setting out priorities and principles for such assessments.
OpenAI said it is outlining four priority areas. It did not name them in the post announcing the commitment, and the paper could not retrieve the published page to check them. The company gave no detail there on who would qualify as an assessor.
The account choblin29 posted that OpenAI says it has already given assessors access to early model checkpoints, visible chain-of-thought and confidential internal data. That detail appears in no other source the paper found, and OpenAI's own post does not state it.
OpenAI named no assessor and gave no timetable. The announcement uses the phrase pacing the frontier, which Anthropic also used on Tuesday in describing Claude Opus 5.5. Neither company links the two statements.