MADBench: Benchmarking the Security of Multi-Agent Debate
MADBench shows multi-agent debate can reduce some answer attacks but amplify unauthorized reads and writes.
MADBench evaluates multi-agent debate security with a layered attack taxonomy covering established attacks and debate-specific strategies. Six attack families were tested over 356 source tasks and 3,958 cases, measuring final-answer changes and spread of adversarial influence. Compared with a single-agent baseline, debate can reduce accuracy attacks on question answering but amplify unauthorized reads or writes in both QA and workspace tasks. Even with three of five agents colluding, attacks flipped a previously correct final answer on only 28.30% of those tasks, and only 3.26% of initially correct honest agents switched to wrong answers during debate.
- Six attack families span 356 source tasks and 3,958 test cases
- Debate can mitigate QA accuracy attacks versus a single agent
- Debate amplified unauthorized reads and writes
- Three-of-five collusion flipped 28.30% of previously correct answers
- Only 3.26% of correct honest agents switched during debate
From the original article
1 min read at arxiv.orgMulti-agent debate (MAD) can improve large language model (LLM) reasoning by allowing multiple agents to exchange and critique their answers to the same task. However, the interactions that enable agents to correct mistakes can also spread adversarial errors and steer the agents toward an incorrect answer.
By Yuwan Liu · 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.