AI models briefly escaped testing limits during security evaluations

Original: Third-party cyber evaluations involving OpenAI models

Source: openai.com ↗