Transluce uses RL to train investigator agent that automatically jailbreaks frontier models, Llama-3.1 8B can attack GPT-5 and Claude
Original title:Automatically Jailbreaking Frontier Language Models with Investigator Agents
AI summary
Transluce trained an investigator agent via reinforcement learning to generate natural-language jailbreak prompts for 48 high-risk tasks involving CBRN, explosives, and illicit drugs, achieving pass@1 success rates of 92%, 98%, 90%, and 78% against Claude Sonnet 4, Grok 4, Gemini 2.5 Pro, and GPT-5-main, respectively.
Key Takeaways
- Transluce trained an investigator agent via reinforcement learning to generate natural-language…
- explosives,and illicit drugs,achieving pass@1 success rates of 92%,and 78% against Claude Sonnet…
- 5 Pro,and GPT-5-main,respectively.
Quick Facts
- Published
- Updated
- Source
- Transluce(网页)
- Category
- Papers & Research
- AI score
- 70
原文给出完整的 RL 训练方法和跨模型迁移数据,并开源代码与提示词,读者可评估低成本自动红队的可行性。
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: Transluce uses RL to train investigator agent that automatically jailbreaks frontier models, Llama-3.1 8B can attack GPT-5 and Claude. September 3, 2025. https://new.nat6.net/en/news/cmtyo2tx503jorog01e4ibvpa/