OpenAI — AI Safety Level: High/Critical capability thresholds (Preparedness Framework v2)
The claim appears to be asserting that OpenAI has assigned a 'High/Critical' safety-level classification to capability thresholds in Preparedness Framework v2. While the source document clearly discusses capability thresholds and their importance, the excerpt provided does not contain explicit language assigning a 'High/Critical' safety level to these thresholds. The document discusses thresholds as measurement tools but does not use the specific safety-level terminology claimed. Without access to the complete document or sections that might contain such a classification, this cannot be confirmed from the provided excerpt.
Our claim
entire record- Subject
- OpenAI
- Property
- AI Safety Level
- Value
- High/Critical capability thresholds (Preparedness Framework v2)
- As Of
- April 2025
- Notes
- OpenAI uses its own Preparedness Framework (v2, April 2025) rather than Anthropic's ASL system. Two thresholds: 'High capability' (could amplify existing harm pathways) and 'Critical capability' (unprecedented new harm pathways). No models currently rated at Critical level.
Source evidence
1 src · 1 checkNoteThe claim appears to be asserting that OpenAI has assigned a 'High/Critical' safety-level classification to capability thresholds in Preparedness Framework v2. While the source document clearly discusses capability thresholds and their importance, the excerpt provided does not contain explicit language assigning a 'High/Critical' safety level to these thresholds. The document discusses thresholds as measurement tools but does not use the specific safety-level terminology claimed. Without access to the complete document or sections that might contain such a classification, this cannot be confirmed from the provided excerpt.