AI safety research groups Apollo Research and Transluce each announced on their own websites on November 16, 2026 (US time) that they will pilot "embedded assessments" that place their researchers inside OpenAI to examine its AI models. The effort goes beyond the conventional approach, in which outside evaluators test a model for a short period just before release, and brings outside scrutiny to models still in training or in internal use.
Both groups describe the effort as a first step and have published its terms and limitations. The detailed conditions are laid out in Transluce's announcement (Announcing an Embedded Assessment Pilot with OpenAI).
Apollo Research Puts a Six-Person Team on a Four-Month Review of Subversive Behavior
According to Apollo Research's announcement, a team of six from the group will carry out the evaluation over four months. It will examine four questions:
- Whether training incentivizes AI models to subvert oversight
- Whether evaluations and red-teaming elicit such behavior
- Whether such behavior actually occurs during internal deployment
- Whether reasoning related to such behavior can be detected
For the first question, Apollo Research will receive access to relevant training code and datasets, reinforcement learning rollouts that include reasoning traces and rewards, and the ability to query intermediate checkpoints. Access for the other three questions remains to be agreed. All of the work will be done on OpenAI hardware, and the findings will be published in two public reports.
Transluce Spends Six Weeks Reviewing OpenAI's Own Alignment Evaluations
Transluce's role is to review the alignment evaluations OpenAI uses in post-training. It will check how well those evaluations catch problems such as deception, manipulation, and reward hacking. The pilot runs for six weeks, with Transluce staff working on-site at OpenAI. Access begins in early December.
Transluce will get privileged access to a handful of OpenAI models, including at least one internal-only model. It will also be able to see raw chain-of-thought, select alignment evaluations, a large number of transcripts, and safety case documentation. At the same time, Transluce states plainly that this access is not comparable to employee-level access.
On costs, Transluce and its staff are not accepting payment from OpenAI. OpenAI provides compute, on-site devices and infrastructure, and temporary workspace. A follow-up assessment with broader and deeper access is planned for early 2027, but its details have not been finalized.
Evaluators Control Publication, and Full Reports Go to a Board Committee
The arrangements also set rules for how results are handled. According to Apollo Research, OpenAI may request redaction of confidential information, but it cannot remove the evaluation's verdicts or findings simply because they are unfavorable. If OpenAI withholds material for technical or security reasons, it must disclose that the material exists and give its reasons in writing. Apollo Research also secured the right to publicly explain the conditions and limitations of its work, including what was withheld, which it calls meta-transparency.
In Transluce's case, the group holds editorial control. OpenAI receives the report before release and may request redactions of intellectual property and non-public information. Results concerning internal-only models will not be published. In both evaluations, the complete, unredacted reports go to the Safety and Security Committee of OpenAI's board. Transluce can also bring disagreements with OpenAI directly to that committee.
Why Pre-Release Testing Alone Falls Short
Apollo Research has evaluated OpenAI models before, but much of that work took place after post-training and shortly before public release. The group says this approach is not sufficient for testing behaviors such as "scheming," in which an AI pursues goals while evading oversight. It also says evaluation results must be credible regardless of whether people trust the developer.
Transluce likewise points out that being embedded gives it access to internal information that has generally not been available to outside evaluators.
Our Take
Giving outside evaluators access to models in training and internal-only models marks a step forward for AI safety evaluation. However, access for three of Apollo Research's four questions has yet to be agreed, and Transluce's pilot lasts only six weeks. How much the evaluators were actually able to see will only become clear once both groups publish what they requested, what was approved, and what was declined, as they have committed to do. We expect that record, along with the reports themselves, to determine how credible this effort turns out to be.
Summary
Apollo Research and Transluce have announced pilots of embedded assessments that examine models from inside OpenAI. Apollo Research will spend four months with a six-person team studying subversive behavior from several angles, while Transluce will spend six weeks reviewing OpenAI's own alignment evaluations. Both groups will publish their results, and unredacted reports will go to a committee of OpenAI's board.