On November 13, 2026, OpenAI published a safety announcement detailing agreements with four independent assessment organizations—Apollo Research, METR, Transluce, and Goodfire—for its first phase of embedded assessments. The post establishes the collaboration framework, scoped resource access, and governance protocols.
What Embedded Assessments Cover Across the Four Assessors
OpenAI defines embedded assessments as giving independent experts deep access to internal models and systems over an extended period to evaluate alignment and safeguards—particularly around safety cases and in support of its independent assessment commitments under the White House Accord on Super Intelligence. Depending on evaluation needs, assessors receive access to some of the same internal resources used by OpenAI researchers, including model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and researcher discussions (or paired analysis with OpenAI staff when direct access could expose sensitive IP):
Apollo Research: Conducting a two-phase pilot evaluating whether models could covertly work against developer intentions (subversion) and whether safeguards detect such behavior across four claims (training incentives, elicitation under evaluations/red teaming, internal-use behavior, and detectability of reasoning), with phase-one access to select public and unreleased models, checkpoints, training/evaluation data, intervention logs, monitoring systems, and compute.
Goodfire: Embedding interpretability researchers using the Goodfire platform to analyze model internals for training-run monitoring, debugging, and alignment—primarily serving OpenAI internal safety researchers while also publishing learnings and findings.
METR: Running an initial alignment assessment focused on misalignment in OpenAI's largest frontier reinforcement learning run to date (narrower than METR's broader alignment survey proposal), with access to checkpoint samples, rollouts, and misalignment monitoring records. Following a three-week monitor red-teaming pilot published October 22, 2026, both parties also tentatively agreed to a broader monitor assessment pending mutual agreement on scope.
Transluce: Conducting an initial evaluation of the quality and coverage of post-training alignment evaluations (with access to select public/unreleased models, evaluation results, behavioral data, safety case documentation, and compute), alongside a shared intent to conduct a roughly eight-week follow-up assessment in early 2027 examining how evaluations inform deployment decisions.
On-Site Security, Oversight & Disclosure Protocols
On-Site Access & SSC Oversight: Assessors handling sensitive confidential information must work on-site at OpenAI offices using OpenAI devices under employee-level (or stricter) security standards. The OpenAI Foundation's independent Safety and Security Committee (SSC) provides oversight before, during, and after assessments, receives unredacted reports, and handles unresolved escalations.
Off-the-Record Discussions, Attestations & Publication: Employee discussions are off the record by default and cannot be attributed to individuals; assessors can request non-attributed attested statements from OpenAI and publish them (or disclose a refusal to attest). Assessors retain publication independence, subject to redaction of confidential details and responsible disclosure delays to remediate risks.
Compensation Disclosure & Cool-Off Rules: Per OpenAI's footnotes, Apollo Research and Goodfire receive financial compensation independent of assessment outcomes, while METR and Transluce receive no financial compensation (accepting only scoped API credits/compute). Assessors are also subject to a brief cool-off period before embedding with another AI developer.
