OpenAI details deep access for safety evaluators; researcher flags oversight gap on misuse

tomekkorbak · x · 2026-11-17

OpenAI safety team member @lamaahmad detailed how third-party evaluators get deep access to some of the most sensitive frontier tools and models researchers use internally, with the process designed to adapt as they learn — aimed at supporting real safety research rather than a checkbox public report. Authorizing employee groups to work with evaluators is not an attempt to restrict conversations, she says.

Researcher tomekkorbak pushed back: the problem with being "precise about what we're doing" is that OpenAI has strong incentives to apply that precision to cut off unfavorable evidence, with no oversight over how this precision is misused.

Related event: Boaz Barak clarifies 'embedded assessors' program: frontier model access, not box-checking(2 posts)→

Original post →

More from Companies & People

Companies & People channel →