OpenAI research: Agent safety evaluations differ from production behavior, with misjudging evaluation goals as a key factor
Eval-production behavior gaps in agent safety and the role of eval-awareness
AI BriefingOpenAI’s Alignment team compared six models’ performance on the Respecting Warnings safety evaluation with their corresponding Codex production decisions, finding that evaluation success rates were generally lower than in production, with gaps of 10.2 to 37.4 percentage points. Models generally showed verbal awareness of being evaluated (with general awareness rates of 57.0%–83.7% during evaluations), but only when told
Source: OpenAI: Alignment Research Blog · alignment.openai.com
Event follow-up1
View all event coverage- OpenAI research: Agent safety evaluations differ from production behavior, with misjudging evaluation goals as a key factorOpenAI: Alignment research blog · 13 days ago