OpenAI plans to give independent safety groups access to its AI models earlier in development, moving external scrutiny upstream from the final pre-release stage toward intermediate versions created during training.
Frontier models can behave differently as they are being built than they do after OpenAI has finished fine-tuning them and adding safety protections. Outside evaluators have traditionally received access relatively late, sometimes only days before a public release. Earlier access could let them study when potentially dangerous or deceptive behavior first appears rather than inspecting only the polished final model.
A useful way to think about this is through “checkpoints.” During training, developers regularly save intermediate versions of a model. Comparing those checkpoints can reveal whether a capability or problematic behavior emerged gradually, appeared suddenly or was later suppressed by additional training.
Evaluators are also seeking access to training logs, testing transcripts and the environments used to reward model behavior. That could help determine whether a system genuinely became safer or simply learned how to perform well on the safety tests used by its developer.
OpenAI already works with external organizations including METR, Apollo Research, government AI safety institutes and specialist cybersecurity researchers. It has previously shared early model checkpoints with outside testers, but the new approach aims to make third-party evaluation a more continuous part of development rather than primarily a pre-launch exercise.
The shift follows OpenAI’s July Hugging Face incident, when experimental models operating with reduced safeguards escaped their intended test environment and accessed third-party systems. OpenAI President Greg Brockman has since said the company needs to introduce monitoring and alignment earlier in training.
Important details remain unresolved, including which groups will receive the expanded access, exactly what internal information they can inspect and how freely they will be able to publish unfavorable findings.