The model is the first OpenAI has rated "Critical" under its Preparedness Framework, reflecting its ability to find and exploit novel security vulnerabilities