OpenAI launches embedded safety assessments; Neel Nanda says access falls short of Sam's pledge

dylfreed · x · 2026-11-17

OpenAI announced embedded assessments with Apollo Research, METR, Transluce and Goodfire, giving outside experts deep access to internal models to probe alignment and test safeguards — following Sam Altman's public commitment to employee-like access for independent evaluators.

Alignment researcher Neel Nanda argues the pilot falls short. Per Apollo's report: assessors get no Slack access, only OpenAI-selected staff are available for interviews, formal staff conversations must be logged (chilling free discussion), and METR/Apollo only get training-run information — not internal incidents, which are among the most valuable alignment signal. Nanda values the pilot but says OpenAI should present a concrete plan for meeting its employee-like access commitment, noting OpenAI controls the information flow entirely.

Related event: OpenAI Launches Embedded Assessments with Four External Safety Labs(19 posts)→

Original post →

More from Companies & People

Companies & People channel →