Skill-Conditional Trust Improves Agent Routing, But Poses Attack Risks.

In platforms that route tasks among diverse LLM agents, an agent's proficiency often varies significantly across different skills. Traditional global trust scores fail to capture this specialization, leading to suboptimal task routing. This paper explores "skill-conditional trust," which assigns a specific trust score to an agent for a particular skill. The study identifies a specific regime where conditional trust offers advantages: when agent heterogeneity is high, per-skill evidence is sparse, and skills are correlated. In such scenarios, borrowing evidence across skills can improve data efficiency. However, this cross-skill borrowing introduces a dual-use vulnerability. The research demonstrates that an attacker can exploit this mechanism. By generating cheap evidence in one skill, an attacker can manipulate the conditional router to misroute tasks in a target skill where they have no direct evidence, leading to significant routing regret. While a zero-evidence gate can mitigate the attack, it doesn't eliminate the residual cost, highlighting a critical trade-off between data efficiency and security in agent swarm management.