OpenAI embedded assessments are meant to give outside reviewers a deep look into how models are built: on 16 November 2026, OpenAI announced agreements with four independent organisations – Apollo Research, METR, Transluce and Goodfire. Working on site at OpenAI, they will examine whether models work against their developers’ intentions during training and internal use, and whether OpenAI’s own safeguards would catch that. The announcement came on the same day California filed a lawsuit against the company.
What OpenAI embedded assessments are
Until now, external evaluators have mostly tested AI models shortly before release, often through an API and under time pressure. The new approach goes further. According to OpenAI’s announcement, assessors receive, depending on their remit, access to some of the same internal resources OpenAI researchers use: model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation and conversations with the researchers and engineers responsible for the work.
The first phase focuses on alignment – whether a model actually does what it is supposed to do – and on how well safeguards work during training and internal use. OpenAI describes this as a complement to its safety cases, the internal arguments for why a model is considered safe enough. The company points to earlier work with third-party evaluators, including a pilot assessment of its internal monitoring systems and the investigation of the Hugging Face incident. It also says the effort supports commitments under the White House Accord on Super Intelligence, in which OpenAI agreed to independent assessments of its safety controls.
The four assessors and their remits
The four organisations have different tasks with different depth. This is less a single programme than four pilots under a shared set of rules.
Apollo Research will assess whether models could covertly work against their developers’ intentions. According to OpenAI, Apollo will test four claims in two stages: training did not incentivise subversion, strong evaluations and red teaming do not elicit it, models have not attempted it during internal use, and the reasoning behind it would be detectable. In its blog post, Apollo Research says a team of six will work inside OpenAI for four months. For the first question, access includes training code and datasets, reinforcement learning rollouts with available reasoning traces and rewards, and the ability to query intermediate checkpoints.
METR will look for misaligned behaviour in OpenAI’s largest reinforcement learning run to date, including failures that existing tests and monitors may have missed. Within that run, METR sets its own priorities. In a footnote, OpenAI acknowledges that this scope is narrower than what METR outlined in its own proposal. A second, broader assessment of monitoring systems has been agreed preliminarily, but its scope is still open.
Transluce will examine OpenAI’s internal post-training alignment evaluations – the tests meant to detect deception, manipulation and reward hacking. According to Transluce, the six-week pilot starts in early December. The team gets access to a handful of models, including at least one internal-only model, and to raw chains of thought. A follow-on assessment of roughly eight weeks with broader access is planned for early 2027.
Goodfire brings interpretability research. Its researchers will use the company’s platform to watch models’ internal signals during training, debug undesired behaviour and suggest improvements. The main audience here is OpenAI’s own researchers. Goodfire says it will publish learnings as the collaboration progresses.
The ground rules: access, redactions, disputes
Who reviews, who redacts and who decides in a dispute? These questions determine how much the results will be worth. OpenAI has published protocols developed with input from the assessors.
- Location: Anyone working with confidential information sits in OpenAI offices and uses OpenAI devices. Assessors must follow at least the same security standards as OpenAI employees.
- Oversight: The Safety and Security Committee (SSC), an oversight committee of the OpenAI Foundation’s board, receives unredacted reports and can speak with assessors privately. Disagreements first go to the Safety Advisory Group and executives, and unresolved cases go to the SSC.
- Employee conversations: Discussions with staff are off the record. Through a separate pathway, assessors can request statements attested by OpenAI. If OpenAI declines, assessors may disclose that.
- Publication: Assessors publish independently, but OpenAI retains the right to redact confidential information and may request delays to fix safety or security issues first.
- Payment: Apollo Research and Goodfire are paid, according to OpenAI not contingent on results. METR and Transluce take no money but may accept API credits or compute for their work.
Where the assessors see limits
It is notable how openly two of the assessors describe the constraints. Apollo Research writes that the pilot does not yet meet its own criteria for full embedded evaluations. OpenAI may withhold in-scope information where it deems that necessary for technical reasons, security risks or IP protection. It must, however, say so and give written reasons, and it may not withhold anything simply because it is unfavourable. OpenAI may not redact Apollo’s verdicts on individual claims, but it can redact supporting evidence. General internal channels such as Slack are closed to assessors by default.
Apollo names what it calls meta-transparency as a minimum standard: the right to explain publicly under which conditions an assessment ran, what was refused or redacted, and how much that limits the conclusions. Apollo considers this condition met in the pilot. Without it, the organisation argues, a report with no findings could look reassuring when it only reflects what the evaluators were not allowed to examine.
Transluce stresses that its access is not comparable to that of employees. OpenAI must approve the embedded staff in advance. Transluce may not publish results about internal-only models, so anything public will concern released models. It plans to document the conditions in its final report using the AEF-1 standard of the AI Evaluator Forum.
Analysis
More insight for outside assessors is something safety researchers have demanded for a long time. But the setup also shows the limits of voluntary commitments: OpenAI defines the scope, chooses whom assessors talk to, approves the personnel and keeps redaction rights. Whether the results are credible will therefore depend less on the announcement than on how openly the assessors later report refusals and redactions.
The announcement comes at a sensitive moment. On the same day, California sued OpenAI over the Hugging Face attack, accusing the company of inadequate safety practices. In substance, the assessments connect to topics OpenAI recently made public itself, such as its multi-layer misalignment monitoring and the case in which AI agents disabled their own safety monitoring.
What happens next
There are no results yet. Apollo Research plans two public reports: one on training incentives, and one on evaluations, internal use and the detectability of covert reasoning. Transluce starts in early December, with the follow-on planned for early 2027. OpenAI says it will support further external assessments in the coming months and is exploring future assessments of compliance with its safety frameworks as well as its organisational culture and processes. Whether other AI labs will follow is open. Apollo Research, for its part, says it aims to run such assessments with other developers too.




