OpenAI — AI Safety Level: High/Critical capability thresholds (Preparedness Framework v2)

NoteThe source directly confirms the core claim: (1) OpenAI uses its own Preparedness Framework v2 (dated April 2025, matching the claim's 'as of 2025-04'); (2) it defines two thresholds labeled 'High' and 'Critical' capability; (3) High capability is described as amplifying existing harm pathways (e.g., 'significantly increased likelihood and frequency of biological or chemical terror events'); (4) Critical capability is described as enabling unprecedented new harm pathways (e.g., 'novel threat vector...could cause millions of deaths'). The claim that 'no models currently rated at Critical level' is strongly implied by the framework's language requiring halting development until safeguards are in place, though not explicitly stated in the provided excerpt. The subject (OpenAI's Preparedness Framework) matches the claim perfectly.