When Should Agent Trust Be Conditional? Characterizing and Attacking Skill-Conditional Reputation in Agent Swarms

The paper introduces the concept of skill-conditional trust (R(i | k)) for heterogeneous agent swarms, addressing the limitations of a single global trust score in task routing. It finds that conditional trust is beneficial in environments with high agent heterogeneity and sparse evidence, providing a small performance gain on a benchmark of 14 heterogeneous agents. However, the study also reveals vulnerabilities, where attackers can exploit conditional routing to significantly increase routing regret, highlighting the need for careful management of trust metrics in AI systems.