OpenAI signs first embedded assessment agreements giving Apollo, METR, Transluce and Goodfire deep internal access

OpenAI · Company announcement

Read the source (opens in a new tab)
What happened
OpenAI said on November 16, 2026 that it signed agreements with Apollo Research, METR, Transluce and Goodfire for a first phase of embedded assessments of its models and safeguards.
Key points
  • Assessors may get checkpoints, source code, training and evaluation data, monitoring systems and researcher interviews; the first work covers alignment and safeguards in training and internal use.
  • Apollo will test whether models could covertly work against their developers' intentions; METR will study misalignment in OpenAI's "largest reinforcement learning run to date."
  • OpenAI's Safety and Security Committee receives unredacted reports and settles unresolved disputes; employee interviews are off the record, with an attestation route for statements assessors may publish.
  • Assessors publish independently, but OpenAI keeps the right to redact confidential information and may request disclosure delays to fix safety or security issues first.
Evidence
OpenAI's own announcement of the agreements and protocols; no assessment under these agreements has reported yet. OpenAI says the work supports its White House Accord commitment to independent assessments of its safety controls.
Uncertain
OpenAI says METR's first-phase scope "is less expansive than what METR outlines" in its own proposal, and a second METR monitor assessment is pending agreement on scope.

See it in the latest record