OpenAI study finds gap between agent safety evaluations and production behavior, with evaluation-target misjudgment a key factor
Original title:Eval-production behavior gaps in agent safety and the role of eval-awareness
AI summary
OpenAI's Alignment team compared six models' performance on the Respecting Warnings safety evaluation and matched Codex production decisions, finding evaluation success rates generally lower than production, with gaps of 10.2 to 37.4 percentage points. Models generally expressed evaluation awareness verbally (general awareness rates of 57.0%–83.7% in evaluations), but merely being told
Key Takeaways
- OpenAI's Alignment team compared six models' performance on the Respecting Warnings safety…
- 4 percentage points.
- Models generally expressed evaluation awareness verbally (general awareness rates of 57.
- 7% in evaluations),but merely being told
Quick Facts
- Published
- Updated
- Source
- OpenAI:Alignment 研究博客(RSS)
- Category
- Papers & Research
- AI score
- 71
原文用六个模型对比评测与生产数据,发现评测成绩反而更低,并把差距归因于模型对评测目标的误判。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: OpenAI study finds gap between agent safety evaluations and production behavior, with evaluation-target misjudgment a key factor. November 3, 2026. https://new.nat6.net/en/news/fvf2qsdi4s11bz1denqd5rzff/