OpenAI Expands Outside Safety Reviews Into Model Training
OpenAI is opening its model training process to outside safety audits — a shift that could reshape how frontier labs validate high-risk systems.
Why it matters
OpenAI is expanding independent safety reviews from post-deployment into the training and evaluation phases themselves, giving external groups earlier visibility into frontier model development. This is a structural change in how capability and safety are validated — relevant to practitioners evaluating model safety claims and to enthusiasts tracking lab accountability practices.
The key facts
7 to knowOutside safety groups granted earlier access to test high-risk systems
Expansion targets model training and evaluation phases, not just post-release
Shift in frontier lab transparency and third-party validation practices
Published September 2026
Independent safety assessments now include model training and evaluation phases
Outside groups gaining earlier access to high-risk AI systems during development
Represents expansion beyond post-release safety reviews
The story so far
Earlier coverage of this storyline
Go to the source
TechRepublictechrepublic.com
Publisher excerpt: OpenAI plans to expand independent safety assessments into model training and evaluation, giving outside groups earlier access to test high-risk AI systems.