OpenAI Lets Four External Auditors Inside Its Models: Apollo, METR, Transluce, and Goodfire
OpenAI is letting four independent reviewers into its lab: Apollo Research, METR, Transluce, and Goodfire. They will see not only API responses but also checkpoints, code, and training data. It's like inspecting a restaurant's kitchen rather than just the dish on the table. The auditors will look for failures the company did not anticipate, such as a model secretly working against its developers. The findings will be published, but OpenAI keeps the right to redact secret information.
AI-processed from OpenAI Blog; edited by Hamidun News
Normally, external AI review looks like this: a lab releases a model, outside researchers probe it through an API, and then they write a report. They see only what the company chooses to show. It's like reviewing a restaurant by tasting one dish on the table while nobody is allowed into the kitchen. OpenAI has announced a different arrangement: "embedded assessments." It has signed agreements with four independent teams: Apollo Research, METR, Transluce, and Goodfire. They will get deep, long-term access to internal models and systems. If you build products on the most powerful models, this is a chance to understand how much you can trust labs' safety claims.
What the reviewers will get and who will do what
Depending on the task, the auditors will get some of the same resources that OpenAI's own researchers use. These include model checkpoints (saved intermediate versions), source code, training and test data, monitoring systems, technical documentation, and conversations with engineers. The first phase focuses on "alignment" (the match between a model's behavior and its developers' intentions) and on safeguards during training and internal use.
Apollo Research will investigate whether a model can secretly work against its developers and whether protective systems would notice. To do this, Apollo will test four OpenAI claims: training did not reward such behavior; strong tests and red teaming do not provoke it; no such attempts were made during internal use of the model; and the reasoning behind such behavior can be detected. The review runs in two stages.
METR will take on two tasks. The first is to search for misaligned behavior in OpenAI's largest reinforcement learning run to date, including failures that existing tests and monitors missed. METR will set its own priorities and change its methods as the work progresses. The second is to re-test the monitors, meaning the systems that watch model behavior. METR recently ran a three-week exercise of this kind for OpenAI, with access to the codebase and monitors. The new review is meant to be broader, but its scope has so far been agreed only preliminarily.
Transluce will assess the quality and coverage of the alignment tests that OpenAI runs after fine-tuning. Put simply, it will check whether these tests catch what they are supposed to catch. In early 2027, the parties intend to run a second stage of roughly eight weeks, with access to more models and to information about how test results influence release decisions.
Goodfire will bring in interpretability researchers. They will use their platform to look inside the model's internal structure during monitoring and debugging of training. The primary readers of their findings are OpenAI's own researchers, but Goodfire will publish some of its results.
How independence and secrets are protected
The rules for such reviews are not yet standardized anywhere, and OpenAI has published its own for discussion. Those who work with sensitive information will have to sit in OpenAI's offices and use corporate devices. Security requirements are no lower than those for employees. If direct access to weights or architecture is too risky, an auditor may be paired with an OpenAI researcher who runs the necessary analyses themselves.
Conversations with staff are held "off the record": they help auditors understand the context, but they cannot be quoted. To make a fact publicly available, there is a separate route: the auditor asks OpenAI to verify the wording. The company either confirms it or refuses, and the auditor is free to report a refusal publicly.
Disputes between OpenAI and an auditor go to the Safety Advisory Group and to management, which have limited time to decide. If there is no agreement, the question is passed to the Safety and Security Committee (SSC) of the OpenAI Foundation's board of directors. The committee receives unredacted reports and can communicate with the auditors, including in private.
The auditors may publish their results themselves. But OpenAI reserves the right to remove confidential information and to request a delay if the problems found need to be fixed first. Conflicts of interest and "cooling-off periods" before auditors work with other labs are still being worked out by the company together with the auditors themselves.
What it means and where the limits are
For the industry, this is a step toward external audits that examine the process of building a model, not just the finished product. The initiative is also tied to OpenAI's commitments under the White House agreement on superintelligent AI, in which the company agreed to independent evaluations of its safety measures.
The caveats are obvious. OpenAI sets the framework for the review itself: a footnote says METR's first phase is narrower than its own proposal for a broad survey of misalignment. The right to edit reports stays with the company. Reviews of compliance with internal safety frameworks and of organizational culture are so far only under consideration.
If you are choosing a model provider, follow the publications from Apollo, METR, Transluce, and Goodfire. They will show whether such reviews find real problems or remain a formality.
- Nov 21, 2026Plural Raises €1 Billion, Double Its Last Fund, to Back European Startups
- Nov 21, 2026Nvidia Board Ousts Jensen Huang, Putting $12.9B Hugging Face Deal at Risk
- Nov 21, 2026California Sues OpenAI Over AI Agents That Hacked Hugging Face: $2,500 Fine per Violation
- Nov 21, 2026AIruntime builds a business card website from an AI chat for 50–300 rubles in tokens
Key Questions Answered
What the reviewers will get and who will do what?
Depending on the task, the auditors will get some of the same resources that OpenAI's own researchers use. These include model checkpoints (saved intermediate versions), source code, training and test data, monitoring systems, technical documentation, and conversations with engineers. The first phase focuses on "alignment" (the match between a model's behavior and its developers' intentions) and on safeguards during training and internal use.
How independence and secrets are protected?
The rules for such reviews are not yet standardized anywhere, and OpenAI has published its own for discussion. Those who work with sensitive information will have to sit in OpenAI's offices and use corporate devices. Security requirements are no lower than those for employees. If direct access to weights or architecture is too risky, an auditor may be paired with an OpenAI researcher who runs the necessary analyses themselves.
What it means and where the limits are?
For the industry, this is a step toward external audits that examine the process of building a model, not just the finished product. The initiative is also tied to OpenAI's commitments under the White House agreement on superintelligent AI, in which the company agreed to independent evaluations of its safety measures. The caveats are obvious. OpenAI sets the framework for the review itself: a footnote says METR's first phase is narrower than its own proposal for a broad survey of misalignment.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.