A man with short, tousled dark hair, wearing a gray crew-neck sweater and a small headset microphone, sits on a stage in front of a blurred yellow and green backdrop.
Sam Altman, OpenAI's chief executive, speaking at TED on April 11, 2025. Archive photograph by Steve Jurvetson (Wikimedia Commons crop, CC BY 2.0); OpenAI announced its first phase of embedded assessments on November 16, 2026. Steve Jurvetson / Wikimedia Commons · CC BY 2.0

What OpenAI Announced

On November 16, 2026, OpenAI said it had signed agreements with four independent assessors, Apollo Research, METR, Transluce, and Goodfire, for a “first phase of embedded assessments.” Depending on what each assessment needs, OpenAI says assessors receive some of the internal resources its researchers use, including model checkpoints, source code, training and evaluation data, monitoring systems, and discussions with staff. Assessors who need sensitive confidential information work on site, on OpenAI devices.

Scopes differ. OpenAI says METR will assess misaligned behavior in its largest reinforcement learning run to date. A footnote calls this scope less expansive than the broad survey METR had proposed, and a broader METR monitor assessment awaits agreement on scope. Apollo says six of its staff will spend four months assessing whether models attempt to subvert their own or their successors’ alignment or control mechanisms. Transluce says its staff will have six weeks of privileged access from early December to assess how well OpenAI’s post-training alignment evaluations capture behaviors like deception and reward hacking. Goodfire says it will use its interpretability platform to monitor training runs.

Redaction, Oversight, and Payment

OpenAI says assessors publish independently, but it retains the right to redact confidential information and may request responsible disclosure delays for findings that could pose security or safety risks. Its Safety and Security Committee, which OpenAI describes as an independent oversight committee of the OpenAI Foundation’s board, receives unredacted reports and settles disagreements that escalation within OpenAI does not resolve. Discussions with staff are off the record. Assessors may publish statements that OpenAI attests to, or report that it declined to attest.

Apollo says OpenAI may not redact its verdict on each claim or request redactions merely because a finding is unfavorable, but may redact evidence it considers sensitive IP. OpenAI may also withhold in-scope items where it deems this reasonably necessary for technical, security, or IP reasons, but must tell Apollo that withheld items exist and give written reasons. Transluce says results on internal-only models cannot be published, and that it will report what it requested, what was approved or declined, and how this affected its conclusions. On payment, OpenAI discloses that Apollo and Goodfire accept financial compensation that is not contingent on their results, while METR and Transluce accept none but may accept API credits or compute. Transluce adds that OpenAI will cover its compute, on-site devices, and workspace.

Limits and Open Questions

Every term here is self-reported, Audit Commons has not seen the agreements, and no findings have been published. As of November 16 we found no post on the agreement on METR’s blog. Apollo says detailed access for its other claims remains to be agreed, and calls the pilot “an ambitious first step towards deeper access and disclosure rights,” though one that does not yet meet the criteria Apollo set out. Goodfire states no duration or redaction terms, and OpenAI describes its main objective as training-run monitoring, debugging, and alignment.

Our editorial reading: read each eventual report beside its stated conditions (access granted, items withheld, redactions, and delays), because those conditions bound what a published conclusion can support.