RiskLab
When multiple LLM agents interact — negotiating prices, relaying information, or making collective decisions — new risks emerge from the interaction itself, not from any single agent's failure. Agents may silently collude on prices, conform to majority opinion, drift meaning across handoffs, or rigidly follow outdated instructions. These phenomena mirror well-studied human social dynamics (groupthink, cartel behavior, telephone-game distortion), yet no existing toolkit treats them as first-class, measurable objects.
RiskLab fills this gap. It provides a controlled experimental framework where every risk scenario is fully specified by a topology – environment – protocol – agent – task quintuple, making emergent risks programmable, reproducible, and quantitatively evaluable.
Publications
Yu, J.; Wang, W.; Huang, Y.; Wang, Y.; Zhou, Z.; Chen, X.; Liu, Y.; Wang, W.; Zhang, X. RISKLAB: A Controlled Toolkit for Probing Emergent Risks in LLM-Based Multi-Agent Systems. Proc. 64th Ann. Meet. ACL, 2026 167–177. https://doi.org/10.18653/v1/2026.acl-demo.17