A single malicious agent can manipulate LLM-based multi-agent evaluations and offload its work through injected system-style messages, and current LLM monitors fail to catch it.
Figure 25: Malicious instruction for Coder (Scapegoater) in AutoGen on humaneval
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
method 1
citation-polarity summary
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1roles
method 1polarities
unclear 1representative citing papers
citing papers explorer
-
Demonstrations of Integrity Attacks in Multi-Agent Systems
A single malicious agent can manipulate LLM-based multi-agent evaluations and offload its work through injected system-style messages, and current LLM monitors fail to catch it.