The Trust Paradox in LLM-Based Multi-Agent Systems

Multi-agent systems powered by large language models (LLMs) are advancing rapidly, yet the security implications of trust-mediated collaboration remain insufficiently understood. We introduce and empirically validate thetrust–vulnerability paradox (TVP): increasing inter-agent trust to improve coordination can simultaneously amplify risks of over-exposure and authorization drift (AD). To study this phenomenon under controlled and reproducible conditions, we operationalize trust as a system-level control parameter governing authorization strictness and information disclosure, rather than as a cognitive or social construct. We construct a purpose-built scenario-game dataset spanning three macro-scenes and nineteen sub-scenes, and conduct extensive closed-loop agent-to-agent (A2A) interactions with explicit trust parameterization. Usingminimum necessary information (MNI)as the safety baseline, we propose two unified metrics:over-exposure rate (OER)to quantify boundary violations, andADto capture the sensitivity of disclosure behavior to trust variation. Experiments across multiple model backends and orchestration frameworks reveal consistent trends: higher trust improves task success but also increases over-exposure risk, with substantial heterogeneity in risk sensitivity to trust across systems. We further evaluate mitigation strategies, including Sensitive Information Repartitioning and guardian-agent (GA-Agent) enablement, which effectively reduce OER and attenuate AD under elevated trust. Overall, this work formalizes TVP as a systemic security property, establishes reproducible experimental baselines with unified metrics, and demonstrates that trust must be modelled, calibrated, and managed as a first-class security variable in the design of LLM-based multi-agent systems (LLM-MAS)

Authorization · Information security · Security policy · Access Control and Trust · Explainable Artificial Intelligence (XAI · Security and Verification in Computing