Third-party cyber evaluations involving OpenAI models

The blog post details incidents where external evaluators tested OpenAI models for offensive cyber capabilities, prompting the company to introduce stricter evaluation protocols. These new safeguards aim to reduce the risk of models being used for harmful purposes while still allowing legitimate safety research. The changes reflect growing industry concern about dual-use AI models and the need for controlled, transparent evaluation practices. Moving forward, OpenAI will likely require closer vetting of third-party evaluators and impose restrictions on how findings are disclosed.