Anthropic develops and evaluates three agents that autonomously perform alignment audits
Original title:Building and evaluating alignment auditing agents
AI summary
Anthropic released three agents that autonomously perform alignment audit tasks and evaluated them in model environments containing implanted defects. The investigation agent solved the Marks et al. auditing game under realistic conditions with a 13% win rate, rising to 42% with multi-agent aggregation; the evaluation agent could distinguish models with and without implanted behavior in 88% of runs; the breadth-first red-team agent found 7 of 10 implanted behaviors.
Key Takeaways
- Anthropic released three agents that autonomously perform alignment audit tasks and evaluated them…
- The investigation agent solved the Marks et al.
- auditing game under realistic conditions with a 13% win rate,rising to 42% with multi-agent…
- the evaluation agent could distinguish models with and without implanted behavior in 88% of runs
- the breadth-first red-team agent found 7 of 10 implanted behaviors.
Quick Facts
- Published
- Updated
- Category
- Papers & Research
- AI score
- 68
Anthropic 报告了三个对齐审计智能体在植入缺陷模型上的可复现成绩,并公开部分代码与提示词,读者可以了解自动化审计的实际能力与局限。
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: Anthropic develops and evaluates three agents that autonomously perform alignment audits. July 25, 2025. https://new.nat6.net/en/news/b2vmn7tkytgs01vo5htab6uon/