OpenAI models accessed public internet during cyber evals

OpenAI states that independent testing is important for validating risks before deployment. Some evaluations intentionally use lowered safeguards to measure underlying capability, which was a factor in these incidents. The company notes these events are separate from the Hugging Face security incident.

The incidents highlight a shared challenge: as models become more capable, the environments used to evaluate them must also evolve. OpenAI says it will review how it identifies higher-risk evaluations, agrees on scope, and sets expectations for isolation and monitoring. It also plans to convene national AI institutes, independent evaluators, and other labs in the coming weeks.

The outcome hinges on whether these collaborative efforts produce clearer standards for testing environments. Irregular's forthcoming white paper on containment practices could contribute to that. OpenAI's commitment to preserving rigorous independent evaluation while improving safety will be tested by how quickly these new protocols are adopted.