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 capability thresholds labeled 'High' and 'Critical' with the exact descriptions provided in the claim ('High capability' = amplifying existing harm pathways; 'Critical capability' = unprecedented new harm pathways); (3) the framework applies to three Tracked Categories (Biological/Chemical, Cybersecurity, AI Self-improvement), each with High and Critical thresholds defined in Table 1. The claim's assertion that 'No models currently rated at Critical level' is consistent with the source's language about safeguards being required before Critical thresholds are crossed, though not explicitly stated in the excerpt provided. The subject (OpenAI's Preparedness Framework v2) matches the claim perfectly.