OpenAI signs agreements with four independent assessors for first phase of embedded safety evaluations
·2026.11.13 17:00
Key point
The initial assessments focus on alignment and safeguards for models during training and internal use, with assessors receiving access to model checkpoints, source code, and monitoring systems.
Details
- OpenAI has signed agreements with Apollo Research, METR, Transluce, and Goodfire for its first phase of embedded assessments, granting independent experts deep access to internal models and systems to evaluate alignment and safeguards.
- Apollo Research will assess four claims regarding subversion, including whether training incentivized it and if reasoning behind subversion is detectable, using access to public and unreleased models and training checkpoints.
- METR will conduct two assessments: one on misaligned behavior in OpenAI's largest reinforcement learning run to date, and a broader monitor assessment pending scope agreement.
- Transluce will examine the quality and coverage of post-training alignment evaluations, with a follow-on assessment of approximately eight weeks planned for early 2027.
- Assessors working with sensitive information will operate on-site at OpenAI offices using provided devices, subject to security standards at least as stringent as those for OpenAI employees.
- The Safety and Security Committee (SSC) will have visibility into the assessments and receive unredacted reports, with disagreements escalated to the Safety Advisory Group and OpenAI executives before final resolution by the SSC.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.