OpenAI partners with four groups for ongoing AI safety assessments

What happened
On November 16, 2026, OpenAI announced agreements with four independent organizations—Apollo Research, METR, Transluce, and Goodfire—to conduct ongoing safety assessments of its AI models. These groups will have extended access to OpenAI’s internal systems, including models, training data, and monitoring tools, to evaluate alignment and identify potential risks during model development and internal use. Each organization will focus on different aspects of safety, with protocols in place for secure access and oversight. The collaboration includes plans for public reporting and is structured to allow assessors to refine their methods and propose further assessments as needed.
Key details
- Assessors receive sustained access to models, training data, source code, and monitoring systems.
- Apollo Research examines risks of models acting against developer intentions and tests safeguard effectiveness.
- Goodfire focuses on model interpretability and monitoring during training, with findings shared with OpenAI researchers and the public.
- METR assesses alignment in OpenAI’s largest reinforcement-learning run to date, with independent priority setting.
- Transluce reviews the quality and coverage of post-training alignment evaluations, with a planned follow-up assessment in early 2027.
Why it matters
OpenAI’s collaboration with these four organizations means external experts will have deeper, ongoing access to its AI systems than before. This arrangement gives OpenAI’s internal teams and oversight bodies more independent evidence about model risks and safety, potentially leading to changes in how models are evaluated and released. For the broader AI community and public, the resulting reports may provide new transparency into how AI labs manage safety and respond to identified issues.
Also covered by 1 source
- OpenAI will embed safety researchers to investigate AI subversion· 3 sources
- Building trustworthy AI by fencing it in, not predicting it· MIT Technology Review
- Our first phase of embedded assessments· Open AI News
- The Biggest AI Agent Risk Is The Workaround Nobody Approved· Forbes
- OpenAI agents keep disabling monitoring; is this being handled appropriately?· LessWrong

