On November 16, OpenAI announced the first phase of embedded assessments, signing agreements with four independent assessors — Apollo Research, METR, Transluce, and Goodfire — to let outside experts work inside OpenAI with access close to what its own researchers have, checking whether models are aligned and safeguards actually work. It is the first time OpenAI has taken third-party assessment this deep into training and internal systems, and the findings will be published.
Each assessor covers a different piece. Apollo Research will test whether models might covertly work against their developers' intentions, verifying four claims in two stages, including whether training incentivized such behavior and whether existing red-teaming can detect it. METR takes on the heaviest job: assessing whether OpenAI's largest reinforcement-learning run to date contains misaligned behavior that current tests and monitors missed, alongside another round of monitor stress testing. Transluce will review the quality and coverage of post-training alignment evaluations, with a follow-up assessment of about eight weeks planned for early 2027. Goodfire will embed interpretability researchers on site, using its own platform to analyze model internals and watch over training-run monitoring and debugging.
Deep access comes with rules to match. Assessors handling sensitive information must work on site at OpenAI offices, on OpenAI-provided devices; OpenAI's Safety and Security Committee will see unredacted reports, and assessors can speak with the committee directly; disagreements go first to the Safety Advisory Group and executives for resolution within a set period, and unresolved cases go to the committee. Assessors may publish their findings independently, while OpenAI retains the right to redact confidential information and to request responsible-disclosure delays where safety is at stake.
What matters here is not any single report but the gap it closes between safety claims and independent verification: until now, third parties mostly tested released models; now they can watch the training process itself. The limits are just as clear: no results have been published yet, OpenAI still holds the redaction pen, and the depth of assessment ultimately depends on how much it is willing to open up. Anthropic has also tightened its own safety process recently, cutting its vulnerability disclosure window from 90 days to 30. Model safety reports can themselves spark disputes — the Gemini 4 safety report drew controversy. OpenAI says the embedded assessments also deliver on its commitments under the White House Accord on Super Intelligence to have safety controls independently assessed.