OpenAI announced agreements on November 16 with Apollo Research, Goodfire, METR and Transluce for the first phase of its program of external assessments embedded in research activity. The projects concern model alignment and safeguards, including during training and internal use. The announcement describes agreements, objectives and access conditions, rather than the final results of the four assessments.
The program aims to give evaluators access to elements beyond ordinary interaction with a public chatbot. Depending on the project, these may include released or unreleased models, intermediate checkpoints, code, training and evaluation data, monitoring tools, documents and discussions with researchers. The scope of access is established separately for each organization involved in the announced assessment program.

What an assessment embedded in research means
The term used by OpenAI describes external evaluators working with the relevant teams and systems inside the laboratory. Access to internal information allows questions different from those put to a model through a public interface. An assessment can, for example, examine how a test is constructed or how training influences the behavior observed during evaluation.
Each engagement begins by agreeing the assessment scope, resources and safety arrangements. Evaluators can follow evidence, refine methods and request additional access as work develops. Where direct access would expose particularly sensitive intellectual property, OpenAI also describes the alternative of analyses conducted together with its researchers. Access is therefore organized around the concrete research question being investigated by the assessor.
Assessors needing sensitive confidential information will work at OpenAI offices on company-provided devices. Security standards applied to experiments will be at least those required of employees, with additional conditions in some cases. Dedicated staff can explain resources and research context while the access arrangements protect the information being consulted during the assessment and associated experimental work.
Four organizations with different assessment objectives
Apollo Research will assess four claims: training did not incentivize subversion; evaluations and adversarial testing do not elicit such behavior; models have not attempted subversion during internal use; and the reasoning behind it would be detectable. These are claims being assessed, rather than conclusions already confirmed by this assessment. The project is organized in two stages.
Goodfire will focus on interpretability and tools for monitoring and debugging training. Interpretability concerns understanding models' internal operation and the relationship between their mechanisms and behaviors. According to OpenAI, the primary audience for this project is initially the internal research team, with lessons from the work to be published as part of the collaboration.
METR will assess alignment within the largest reinforcement-learning training run covered by its agreement. The organization can set independent priorities within this agreed scope. An additional, broader project concerning monitoring remains at a preliminary stage, with the scope of that work still to be defined between the parties rather than already fixed in the announcement.
Transluce will examine the quality and coverage of evaluations in post-training. The company also mentions the intention of an eight-week follow-on collaboration in early 2027. That stage is presented as a future plan; it is neither an assessment already completed nor a public report of findings available at the time of this program announcement.
Scopes of the first phase announced by OpenAI
| Organization | Project focus | Boundary |
|---|---|---|
| Apollo Research | Alignment and detection of problematic behavior | Selected claims and mechanisms for assessment |
| Goodfire | Interpretability and training monitoring | Initial usefulness for internal research |
| METR | Alignment in a major reinforcement-learning run | The additional monitoring project is preliminary |
| Transluce | Quality of post-training evaluations | The eight-week stage in 2027 is planned |
Who receives reports and how disagreements are handled
OpenAI describes an oversight role for the Safety and Security Committee, a committee of the OpenAI Foundation board. Evaluators can communicate directly and privately with it. The committee receives reports without the redactions applied to public versions, separating the information available for oversight from the form in which findings will eventually be communicated outside the program.
Disagreements are escalated to the Safety Advisory Group and OpenAI executives, with a specified resolution period and prompt notification of the committee. If a disagreement remains, the matter is referred to the committee for resolution. This process describes oversight and resolution responsibilities within the program, separately from the technical work through which evaluators produce their findings.
For publication, OpenAI retains the right to redact confidential information and request delays related to responsible disclosure of problems. These conditions form part of the agreements described. The existence of an internal report and the appearance of a complete public version on the same day are therefore different commitments rather than a single program requirement.
Funding and independence have explicit conditions
Apollo Research and Goodfire receive compensation for their work, with payment not contingent on results favorable to OpenAI. This detail concerns the contractual relationship described by the company. Each project's conclusions can be examined through the scope assessed, evidence presented and limitations of the published report, rather than only through the name of the participating organization.
METR and Transluce do not accept financial compensation for assessment, according to the conditions presented. These organizations may, however, receive compute resources or API access. The distinction matters for transparency: receiving no assessment payment and receiving no resources of any kind from the laboratory are different circumstances, which the agreements describe separately within the program.
Employee discussions can inform an assessment but cannot be published or attributed to individuals. Separately, assessors can ask OpenAI to verify, correct or clarify information through attested statements. They can publish statements the company attests to, or disclose that it declined attestation. The procedure separates confidential conversational context from a statement the company is prepared to stand behind publicly.
What the announcement means for business applications
An assessment of a laboratory and an assessment of a business application concern different levels. The first may examine training, alignment or model evaluations. The second includes business data, permissions, connected tools and the outcome of a concrete task. OpenAI's program describes the first level and does not publish an individual audit of customer implementations.
This separation matters for custom AI applications, where the model is one component. An agent reading documents and an agent able to modify records have different permissions and effects, even when they use the same model family. The provider's assessed scope and the application's scope need separate descriptions in the documentation of the implementation.
The development announced on November 16 is the establishment of projects with internal access and explicit objectives for four evaluators. Results, methods and any findings published later will constitute evidence separate from this announcement. At this stage, the concrete information is the organizations involved, assigned topics, access conditions and mechanisms governing oversight and publication of the work.