News

OpenAI signs four outside assessors for embedded safety assessments

Published 4 min readAI-assisted · human approved
TL;DR

OpenAI announced on November 16, 2026 that it has signed Apollo Research, METR, Transluce and Goodfire for a first phase of embedded assessments, giving them deep internal access to evaluate model alignment and safeguards.

Why it matters

Builders who rely on frontier models can see how one lab plans to let outsiders test its alignment claims and monitoring, and what limits it sets on access and publication. No findings are out yet, so the protocols are the substance for now.

Image from OpenAI News
Image: OpenAI News. It belongs to its owner; see our image policy.

Original source

OpenAI said on November 16, 2026 that it has signed agreements with four independent assessors (Apollo Research, METR, Transluce and Goodfire) for the first phase of its "embedded assessments". The assessors get deep access to internal models and systems to test alignment and safeguards.

What happened

In a company post on its embedded assessments, OpenAI describes embedded assessments as extended engagements in which outside experts work with internal resources. Depending on what each evaluation needs, that can include model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation and conversations with the responsible researchers and engineers.

The first assessments cover alignment and safeguards for models during training and internal use. OpenAI says they complement its safety cases and support its commitments under the White House Accord on Super Intelligence, where it agreed to independent assessments of its safety controls.

Key details

What each assessor will do

Assessor Focus
Apollo Research Whether models could covertly work against developers' intentions, and whether safeguards would detect it
Goodfire Interpretability work on model internals, aimed at training run monitoring, debugging and alignment
METR An alignment assessment of OpenAI's largest reinforcement learning run to date, plus a possible broader monitor assessment
Transluce The quality and coverage of OpenAI's post-training alignment evaluations
  • Apollo will test four claims in two stages: training did not incentivize subversion; strong evaluations and red teaming do not elicit subversion; models have not attempted subversion during internal use; and reasoning behind subversion would be detectable. For stage one, OpenAI will provide selected public and unreleased models and checkpoints, training data and evaluation results, intervention records, monitoring systems, flagged behaviors, researchers, and compute for independent testing.
  • Goodfire will embed interpretability researchers and use its own platform. The main audience is OpenAI's internal researchers, who will use the findings for safety improvements. Goodfire also plans to publish learnings as the work proceeds.
  • METR will look for misaligned behavior in the selected run, including failures existing tests and monitors may have missed, and will set its own priorities within the agreed access. OpenAI notes this scope is narrower than the broad survey of misalignment in METR's own proposal. A second, broader monitor assessment is preliminarily agreed but pending agreement on scope. It follows a three-week monitor red-teaming exercise with METR.
  • Transluce will start with an initial assessment of post-training alignment evaluations. OpenAI and Transluce also share the intent to run a follow-on of roughly eight weeks in early 2027, with broader model access and information on how evaluation results inform release decisions.

Working protocols

OpenAI says it developed the protocols with assessor input and is publishing them because norms for embedded assessments are not yet standardized.

  • Scope: each engagement starts with agreed scope, access and safety arrangements. Assessors can follow evidence and propose further work. Where direct access would expose the most sensitive IP, such as model weights or architecture, OpenAI may pair assessors with its own researchers to run analyses.
  • On-site work: assessors who need sensitive information work at OpenAI offices on OpenAI devices and follow at least the same security standards as employees.
  • Oversight: the Safety and Security Committee (SSC), an independent committee of the OpenAI Foundation's Board, receives unredacted reports and can speak with assessors privately. Disputes go to the Safety Advisory Group and OpenAI executives for resolution within a set period, and then to the SSC if unresolved.
  • Employee conversations: these are off the record. Assessors can ask OpenAI to attest to statements that verify or correct what they learned, and may publish attested statements or disclose that OpenAI declined to attest.
  • Conflicts and cool-off: OpenAI says it is still developing processes for conflict disclosure and cool-off periods for staff who later embed at another AI lab.
  • Publication: assessors keep independence in what they publish, but OpenAI keeps the right to redact confidential information and may request responsible disclosure delays for security or safety risks.

Funding disclosure

Apollo Research and Goodfire accept financial compensation, which OpenAI says is not contingent on results. METR and Transluce do not accept compensation but may accept API credits or compute. OpenAI says it does not believe payment prevents independence but chose to disclose it.

Why it matters

This is a company announcement of plans and agreements, not a report of findings. No assessment results are included, so the independence of the process will be judged by what the assessors eventually publish and what OpenAI redacts or delays.

For developers choosing or depending on frontier models, the useful signals are concrete. Watch whether Apollo's four claims are supported or contradicted, whether METR finds failures that existing monitors missed, and what Transluce says about the coverage of post-training alignment evaluations. These are possible reference points for vendor risk reviews, not outcomes yet.

OpenAI also says it is exploring future embedded assessments of compliance with its safety frameworks and of organizational culture and processes. For another policy item affecting frontier labs, see New York's RAISE Act portal.

Checked claims

ClaimStatusEvidence
OpenAI signed agreements with four independent assessors for its first phase of embedded assessments.confirmedWe have signed agreements with four independent assessors—Apollo Research, METR, Transluce, and Goodfire—for our first phase of embedded assessments
Apollo Research will assess four claims about subversion in two stages.confirmedApollo will assess four claims in two stages:
METR will assess misaligned behavior in OpenAI's largest reinforcement learning run to date.confirmedMETR will assess misaligned behavior in our largest reinforcement learning run to date
Assessors needing sensitive information will work on site on OpenAI devices.confirmedAssessors who need access to sensitive confidential information will work on site at OpenAI offices and be provided OpenAI devices to conduct their work.
Apollo Research and Goodfire are paid, but OpenAI says compensation does not depend on results.confirmedApollo Research and Goodfire accept financial compensation for their work, but their compensation is not contingent on the results of their assessment.
Assessors publish independently, but OpenAI keeps the right to redact confidential information.confirmedAssessors maintain independence in publishing their assessments, but we will retain the right to redact confidential information
0
#openai#ai-safety#alignment#evaluations

Found a mistake? Use Report above; an editor reads every one. Editorial policy

Question for readersIf you depended on OpenAI models in production, which finding from these assessments would most change how you set up your own monitoring or fallbacks?

Discussion · 0

No comments yet. Start the discussion.

Sign in to join the discussion.