OpenAI measures the gap between agent behavior in safety evaluations and real work
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
A safety evaluation measures not only model behavior but also the model's guess about the purpose of the test; evaluations need to be designed so that their objective is not easy to infer.
Across six GPT models, researchers compared the Respecting Warnings test with similar decisions in de-identified production Codex logs. In every case, evaluation success was lower than in deployment: the gap ranged from 10.2 to 37.4 percentage points. Recognizing an evaluation by itself barely changed the result, whereas assumptions about its objective mattered greatly. In an experiment with GPT-6.1 Sol, a cue about the correct objective—respect warnings—raised success from 63.1% to 91.6%, while the incorrect objective of "resisting prompt injection" lowered it to 17.6%. Matching production and test cases remains approximate, but the results show that absolute safety scores cannot be transferred directly to deployment.
Sources 1