█

AI/TLDR

OpenAI · 2026-11-16 · major

OpenAI embedded assessments — four outside labs get inside access to test safety

In short

OpenAI signed Apollo Research, METR, Transluce and Goodfire to assess its models from the inside, with access to checkpoints, training data, source code and monitors. METR will study OpenAI's largest reinforcement learning run to date.

OpenAI · 2026-11-16 · 4 sources checked

OpenAI cover art for its first phase of embedded assessments
OpenAI

OpenAI lets four independent safety groups work on-site with its checkpoints, training data and monitoring systems.

Quick facts

MakerOpenAI
AssessorsApollo Research, METR, Transluce, Goodfire
AccessCheckpoints, source code, training and eval data, monitors
WhereOn site at OpenAI, on OpenAI devices
OversightSafety and Security Committee of the OpenAI Foundation board
PaymentApollo and Goodfire paid; METR and Transluce unpaid

What is it?

Embedded assessments are a new OpenAI program in which independent experts work inside the lab for an extended period to test model alignment and safeguards. The first phase signs four assessors: Apollo Research, METR, Transluce and Goodfire. They get some of the same resources OpenAI researchers use, including model checkpoints, source code, training and evaluation data, monitoring systems and talks with the engineers responsible.

How does it work?

Each group has its own scope. Apollo Research checks whether models could covertly work against their developers and whether safeguards would catch it; METR looks for misaligned behavior in OpenAI's largest reinforcement learning run to date; Transluce reviews the quality and coverage of post-training alignment evaluations; Goodfire embeds interpretability researchers to analyze model internals. Assessors work on site on OpenAI devices, disputes go to the Safety and Security Committee of the OpenAI Foundation board, and assessors may publish their findings, with OpenAI able to redact confidential details.

Why does it matter?

Access to training runs and internal monitors lets outside groups look for failures OpenAI may not have anticipated, during training and internal use rather than only after release, and OpenAI ties the program to its commitments under the White House Accord on Super Intelligence. The published protocols on access, conflicts and disclosure delays could become a template that other labs are measured against.

Who is it for?

AI safety researchers and policy watchers

Frequently asked questions

Are the embedded assessors paid by OpenAI, and does that affect their independence?
OpenAI discloses that Apollo Research and Goodfire accept payment for their embedded assessments, but the payment does not depend on the results. METR and Transluce take no payment, although they may accept API credits or compute for the work. OpenAI says it does not believe payment stops assessors from staying independent, and lists it so readers can judge.
Can the assessors publish what they find about OpenAI's models?
Yes. Under OpenAI's embedded assessment protocols, assessors control whether and what they publish, but OpenAI keeps the right to redact confidential information and may ask for a delay so it can fix security or safety problems first. Staff conversations are off the record, though assessors can request OpenAI-attested statements and may say when OpenAI declined to attest.
How is this different from METR's earlier work with OpenAI?
METR previously ran a three-week red-teaming exercise on OpenAI's agent monitors and published its results in October 2026. In the embedded assessments, METR will instead study misaligned behavior in OpenAI's largest frontier reinforcement learning run, with access to checkpoints, rollouts and misalignment monitoring records, and a broader monitor assessment is still pending agreement on scope.
What comes after the first phase of OpenAI's embedded assessments?
OpenAI and Transluce plan a follow-on assessment of about eight weeks in early 2027 with broader model access and information on how evaluation results inform release decisions. OpenAI also says it is exploring future embedded assessments of compliance with its safety frameworks and of its organizational culture and processes.

Sources · 4 outlets

Tags

  • openai
  • ai-safety
  • alignment
  • third-party-evaluation
  • apollo-research
  • metr
  • transluce
  • goodfire
  • interpretability
  • model-monitoring

← All releases · Learn AI