๐ November 16, 2026 โข OpenAI
OpenAI has signed agreements with 4 independent assessors โ Apollo Research, METR, Transluce and Goodfire โ for the first phase of its embedded assessments program. Under this model, outside experts receive deep access to internal models and systems over an extended period to assess model alignment and safeguard effectiveness during training and internal use, not only before release. The announcement post is dated November 16, 2026 on openai.com, and a wave of independent coverage from qz.com, TechCrunch and CNBC followed on November 16, 2026.
Access under the agreements can include model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation and discussions with the researchers and engineers behind the work. Apollo Research will keep 6 evaluators inside OpenAI for 4 months to investigate whether models could covertly work against their developers' intentions and whether safeguards would detect such behavior, drawing on evidence from training, safety testing and internal use. Apollo structures the work as 2 stages covering 4 claims and plans 2 public reports. The lab also secured meta-transparency rights to publicly describe what access it was denied and how redactions affect its conclusions, while calling the pilot an ambitious first step short of its own bar for a full embedded evaluation.
Goodfire embeds interpretability researchers who use the Goodfire platform to analyze model internals, with a focus on training run monitoring, debugging and alignment, publishing learnings as the collaboration progresses. METR is assessing misaligned behavior in OpenAI's largest reinforcement learning run to date, including failures that existing tests and monitors may have missed, and sets assessment priorities independently. A further, broader monitor assessment awaits agreement on scope. Transluce is reviewing the quality and coverage of OpenAI's post-training alignment evaluations, the parties intend a follow-on assessment of about 8 weeks in early 2027, Transluce staff get 6 weeks of privileged access from early December, and results on internal-only models cannot be published.
Several caveats frame the news. No findings have been published yet, since phase one is about access and process. Scopes are negotiated per assessment and in places are narrower than what assessors originally proposed. The announcement also comes from the vendor itself, so independent coverage confirms the agreements while no results exist yet. For oversight, the Safety and Security Committee of the OpenAI Foundation's Board receives unredacted reports, with disagreements escalating to the Safety Advisory Group and OpenAI executives. Apollo Research and Goodfire accept compensation not contingent on results, while METR and Transluce accept no money but may take API credits or compute. Assessors keep independence in publishing, OpenAI may redact confidential information and request responsible disclosure delays, and OpenAI ties the effort to its commitments under the White House Accord on Super Intelligence.
Sources:
openai.com/index/our-first-phase-of-embedded-assessments/
qz.com/openai-four-independent-safety-evaluators-embedded-assessments-111626
aiweekly.co/alerts/openai-embeds-apollo-metr-transluce-goodfire-in-safety-audit
auditcommons.org/news/openai-embedded-assessments-first-phase
www.apolloresearch.ai/blog/pilot-towards-embedded-evaluations
Continue the conversation: https://t.me/AIAutomationArchitecture ยท Bluesky
Top comments (0)