ReMAS red-teams multi-agent systems to extract system prompts
Original title: Black-Box Red-Teaming of Multi-Agent Systems via Reinforcement Learning
A rewrite stage and attack templates raise success versus prior black-box methods and transfer across backends.
A paper introduces ReMAS, a reinforcement-learning framework that fine-tunes attacker language models to extract system prompts from multi-agent systems. It first rewrites base prompts, then builds templates meant to invoke specific sub-agents. The authors report higher attack success than existing black-box methods, with transfer across backend models and system structures. They argue current single-agent attacks are much less effective when sub-agents do not talk directly to the user.
Unconfirmed: Attack-success gains are the authors' experimental claims; independent reproduction is not reported.
- The stated goal is system-prompt extraction, not a deployed incident.
- Experiments claim transferability across different backend LLMs and MAS structures.
- Prior black-box red-teaming is described as far less effective against multi-agent setups.
Impact: extraction of system prompts from multi-agent LLM systems
From the original article
1 min read at arxiv.orgLarge language model (LLM) agents are increasingly deployed in multi-agent systems (MAS) to accomplish complex tasks. Prior black-box red-teaming attacks mainly focus on single agents, but we find that these methods are far less effective in MAS, where multiple sub-agents may not directly interact with the user.
By Yuqi Jia · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.