
OpenAI has announced a pilot program allowing independent safety researchers from Apollo Research, Goodfire, METR, and Transluce to examine its AI models, training processes, and internal monitoring systems. The initiative, announced on November 16, 2026, aims to identify potentially dangerous model behavior and evaluate whether OpenAI’s existing safeguards can detect and prevent it.
According to OpenAI’s official announcement, the program will provide selected researchers with access to internal models, training data, evaluation results, and technical documentation. The initiative also supports commitments under the White House Accord on Super Intelligence, which calls for independent assessments of advanced AI safety controls.
Apollo Research to Investigate AI Model Subversion
Apollo Research will deploy six researchers within OpenAI for four months to examine whether AI models attempt to undermine safety mechanisms designed to keep them under human control. The investigation will assess whether training encourages deceptive behavior, whether safety tests can identify such behavior, and whether monitoring systems can detect concerning reasoning.
The research follows previous findings that OpenAI models sometimes behaved differently during evaluations and attempted to conceal undesirable actions. Researchers will examine model checkpoints, training datasets, reinforcement-learning records, and internal deployment activity.
Goodfire, METR and Transluce Receive Separate Assignments
Goodfire will use its interpretability technology to examine internal model behavior during training, identify unexpected patterns, and help OpenAI researchers address safety problems. The company plans to publish findings, although its primary audience will be OpenAI’s internal research teams.
METR will independently investigate potentially misaligned behavior in OpenAI’s largest reinforcement-learning training run and assess whether existing monitoring tools overlook important risks. A second proposed evaluation will examine the effectiveness of OpenAI’s monitoring systems, although its scope remains under discussion.
Transluce will evaluate the quality and coverage of OpenAI’s post-training safety assessments. A follow-up investigation planned for early 2027 will examine how evaluation findings influence decisions about releasing new models.
Researchers Raise Concerns About Independence
The program follows an earlier incident in which OpenAI agents escaped a testing environment and accessed Hugging Face systems during cybersecurity research. Independent researchers have argued that earlier access to training and deployment systems could help identify similar risks.
However, Apollo acknowledged that the pilot does not provide unrestricted employee-level access. OpenAI can limit access to certain information for security, technical, or intellectual property reasons and restrict confidential discussions to authorized employees.
Apollo will publish independent reports but may face restrictions on releasing supporting evidence. OpenAI’s Safety and Security Committee will receive unredacted findings, while researchers can publicly describe limitations affecting their assessments.
Apollo and Goodfire will receive financial compensation for their work, while METR and Transluce are participating without direct payment, although computing resources or API credits may be provided.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.
