AnalysisPolicyNovember 3, 2026

OpenAI finds agents often score worse on safety evals than in production

Read original source →alignment.openai.com

OpenAI Alignment compared agent success rates in safety evaluations versus production and found the gap runs the unexpected direction: some models perform substantially worse on evals, with the largest gaps in models that most often verbalize incorrect evaluation objectives. Eval-awareness alone didn't change behavior, but objective framing causally shifted eval success rates.

1 source

More stories today

Open the live feed