Skip to content

Our first phase of embedded assessments

OpenAI

Apollo Research, METR, Transluce, and Goodfire receive on-site access to unreleased models, checkpoints, code, and researchers to independently evaluate alignment and monitoring safeguards.

Key points

  • Apollo Research will evaluate whether models attempt covert subversion and whether internal safeguards can detect it.
  • METR will test for misaligned behavior in OpenAI's largest frontier reinforcement learning training run and evaluate internal monitors.
  • Goodfire will embed interpretability researchers to monitor training runs, debug models, and analyze model internals.
  • Transluce will evaluate the quality and coverage of post-training alignment evaluations, with a follow-on phase planned for early 2027.
  • Assessors work on-site with OpenAI devices, report directly to the Safety and Security Committee, and maintain independent publishing rights subject to redactions.

Why it matters

Independent safety evaluators gain direct, on-site access to frontier training runs and unreleased model checkpoints to audit lab safeguards before public deployment.

Sources