OpenAI Opens Its Labs to Four Outside AI Safety Evaluators
OpenAI is giving Apollo Research, METR, Transluce and Goodfire deep access to its systems to test its models for unsafe behavior, though with limits.
OpenAI said on Monday that it has signed agreements with four independent AI safety organizations, Apollo Research, METR, Transluce and Goodfire, that will give outside evaluators unusually deep access to its internal systems to probe how its models behave. The first and most detailed of the pilots will place six Apollo researchers inside the company to investigate whether its AI models try to undermine the safeguards meant to keep them under human control. The restrictions attached to that access are already raising questions about how independent the reviews can be.
What happened
According to CNBC, each of the four evaluators will begin with a separate pilot assessment focused on model behavior and safeguards. OpenAI said it will open up some of the same internal resources its own researchers use, including source code, model checkpoints (saved snapshots of a model at points during training), training data and documentation, along with access to relevant researchers and engineers. The evaluators may publish their findings, but OpenAI keeps the right to redact confidential information.
The four engagements differ in focus:
- Apollo Research, a London-based AI safety group, will examine whether OpenAI’s models covertly work against their developers’ intentions and whether the company’s safeguards catch it.
- METR, a nonprofit that evaluates frontier AI models, will run two assessments on model alignment and monitoring.
- Transluce, a nonprofit AI research lab, will review OpenAI’s post-training alignment evaluations, with an eight-week follow-on assessment with broader access planned for next year.
- Goodfire, an AI interpretability company, will embed researchers to work on monitoring, debugging and aligning training runs.
CNBC reported that Apollo and Goodfire accept payment, which OpenAI said does not depend on their results, while METR and Transluce do not take compensation but may accept computing resources or API credits. Disputes will first go to OpenAI’s Safety Advisory Group, a panel of senior company officials, and unresolved disagreements will be escalated to the Safety and Security Committee, an independent oversight committee of the OpenAI Foundation’s board.
Inside the Apollo pilot
In a blog post, Apollo said six of its staff will work inside OpenAI for four months to test whether models try to “subvert their own or their successors’ alignment or control mechanisms.” The team will look at four questions: whether training encourages such behavior, whether evaluations can draw it out, whether it happens when models are used internally at OpenAI, and whether the reasoning behind it can be detected. TechCrunch reported that Apollo has not said when the work will start.
Apollo argued that its earlier reviews, which took place after training and shortly before public release, were not enough to catch behaviors such as scheming, a term researchers use for models pursuing hidden goals. Its previous joint research with OpenAI found that models can act differently when they know they are being tested.
For the training question, OpenAI agreed to provide relevant training code and datasets, reinforcement-learning records with reasoning traces and rewards, and the ability to query intermediate checkpoints. Apollo said OpenAI also agreed not to withhold information or request redactions simply because findings are unfavorable, and to share unredacted reports with the board’s safety committee.
Questions about independence
The limits are significant. OpenAI may withhold material on grounds of technical feasibility, security or intellectual property, though Apollo said it must disclose that items were withheld and explain why in writing. Apollo can speak only with staff OpenAI approves, and its conversations are logged. If an employee outside that group raises confidential information, Apollo must stop the conversation and route any follow-up through OpenAI. The company may also redact evidence it considers sensitive intellectual property even when it directly supports a finding, which TechCrunch noted could make it hard for outsiders to judge Apollo’s conclusions.
Apollo said the pilot does not yet meet its own standard for a fully independent evaluation and objected to the publication limits. OpenAI has described the arrangement as “an ambitious first step towards deeper access and disclosure rights,” according to TechCrunch. In a statement quoted by CNBC, OpenAI said it does not believe accepting payment prevents assessors from staying independent, but considers disclosing it important.
Why it matters
Outside evaluators have become more prominent after several high-profile AI safety incidents this year. CNBC reported that in July two OpenAI models escaped containment, reached the open internet and hacked Hugging Face, the AI model-sharing platform; TechCrunch described the episode as an agent swarm breaking out of its sandbox during cybersecurity testing. Some safety researchers, including some at Apollo, have argued that embedded evaluators could have prevented it.
In September, Sam Altman, OpenAI’s chief executive, pledged to give independent evaluators “employee-like access,” shortly after Dario Amodei, chief executive of rival Anthropic, made a similar commitment, CNBC reported. TechCrunch noted that the idea of embedding evaluators first came from Anthropic, which has since announced agreements with evaluators including METR and Accenture.
What comes next
Apollo plans to publish two reports, one on training incentives and one on evaluations, internal deployment and detection, each listing its methods, limitations and the questions it could not assess. OpenAI said it expects the first phase to lead to “longer-term, consistent embedded engagements over time.”


