-+ 0.00%
-+ 0.00%
-+ 0.00%

OpenAI Brings Four Outside Research Groups Into AI Safety Testing

Benzinga·11/16/2026 17:41:24
Listen to the news

In a blog post, OpenAI announced the first phase of its "embedded assessments" program, partnering with Apollo Research, Goodfire, METR and Transluce to conduct extended evaluations of its models and safety protections.

The initiative is designed to give outside researchers a closer look at OpenAI’s internal safety work and produce findings that can eventually be shared publicly.

Under the agreements, researchers may receive access to model checkpoints, source code, training and testing datasets, monitoring infrastructure, technical documentation and direct discussions with the engineers and researchers responsible for developing the systems.

Each Organization Focuses On a Different Area of AI Safety

  • Apollo Research will investigate whether OpenAI models could secretly act against their developers’ intentions and whether the company’s safeguards would detect such behavior. Its assessment will examine training incentives, testing and red-teaming results, internal model behavior and potential signs of subversion.
  • Goodfire will focus on interpretability, using its platform to examine how models work internally. The research will concentrate on monitoring and debugging models during training, with plans to share findings and lessons publicly.
  • METR will pursue two evaluation tracks. One will look for potentially misaligned behavior during what OpenAI described as its largest reinforcement-learning training run to date. A second effort could expand previous research into stress-testing AI monitoring systems, although the scope of that project has not been finalized.
  • Transluce will review OpenAI’s post-training alignment evaluations, examining how comprehensive the tests are and what types of risks they cover. The two organizations also plan to pursue a follow-up project lasting roughly eight weeks in early 2027, with broader access and a focus on how evaluation results influence model-release decisions.

OpenAI Outlines Rules For Independent Assessments

OpenAI said it developed operating protocols with input from the research groups to balance technical access with confidentiality, security and research independence.

The framework includes on-site access to sensitive materials using OpenAI-provided devices, oversight from the OpenAI Foundation board’s Safety and Security Committee, and procedures for resolving disagreements and coordinating publication timelines.

The company also outlined restrictions on conversations with employees. Discussions with OpenAI staff will remain off the record and cannot be published or attributed, although researchers can use a separate process to obtain company-attested statements for verification or clarification.

OpenAI said it will disclose potential conflicts of interest and apply cooling-off periods as part of the program’s independence safeguards.

The financial arrangements also differ across the four groups. Apollo Research and Goodfire will receive compensation that is not tied to the outcomes of their assessments, while METR and Transluce will not receive direct payments but may use credits to cover computing costs.

The first phase aims to give outside researchers deeper access to the development and testing of advanced AI systems, with OpenAI positioning independent evaluations as an additional layer of scrutiny beyond its internal safety processes.

Photo courtesy: Meir Chaimowitz / Shutterstock