openai.com web signal

OpenAI Embeds Apollo, METR, Transluce, Goodfire in Safety Audit

OpenAI Safety ai-safety

TL;DR

  • Apollo Research is placing six evaluators inside OpenAI for four months to probe whether frontier models covertly pursue goals that conflict with developer intent.
  • METR will hunt for misaligned behavior inside what OpenAI calls its largest reinforcement-learning run to date; Goodfire embeds interpretability researchers to study model internals.
  • Apollo and Goodfire accept paid engagements not tied to findings; METR and Transluce decline cash but may accept compute or API credits.

OpenAI has signed agreements with four independent safety labs to embed researchers inside its own training runs, with six evaluators from Apollo Research spending four months examining whether the company's models "covertly pursue goals that conflict with their developers' intentions," OpenAI said in a post describing the first phase of the program.

Apollo's brief is organised around four questions: whether training incentivises subversion, whether evaluations can elicit it, whether it occurs during internal deployment, and whether the relevant reasoning would be detectable. The lab gets training code, datasets, reinforcement-learning records and intermediate model checkpoints. METR, per CNBC, will hunt for misaligned behavior inside what OpenAI describes as "OpenAI's largest reinforcement-learning run to date," looking for failures existing tests and monitors may have missed. Goodfire is embedding interpretability researchers to analyse model internals during training. Transluce is auditing the quality and coverage of OpenAI's post-training alignment evaluations; OpenAI "plans to expand the pilot in 2027."

Compensation terms split the four labs. Apollo and Goodfire take payment, "although compensation does not depend on the assessment's results." METR and Transluce accept no money but may take compute or API credits. OpenAI retains redaction rights over confidential material, yet Apollo says OpenAI "may not redact Apollo Research's verdicts" and must identify any withheld material in writing. Apollo will publish two public reports; the Safety and Security Committee of the OpenAI Foundation's board receives unredacted versions and resolves disputes escalated past the company's Safety Advisory Group.

Goodfire, founded by researchers from OpenAI and Google DeepMind, "has raised more than $200 million, most recently at a $1.25 billion valuation." The embedded-assessments announcement lands the same day OpenAI rolled out a CSAM image classifier with NCMEC, part of a visible cluster of safety-adjacent disclosures from the company this week.