RESEARCH · GLOBAL
arXiv: Trustworthy Agent Networks—trust must be architected, not bolted on
Position paper argues that multi-agent LLM systems (agents coordinating to solve tasks) introduce cascading failures, adversarial composition, and semantic misalignment. Trust must be baked into design, not added post-hoc. Existing alignment techniques insufficient for A2A networks.
WHY IT MATTERS
As banks deploy multi-agent agentic workflows (Woori Bank's 29 ops), security and compliance architectures must account for agent-to-agent failure modes; CROs and Chief Information Security Officers should mandate agent audit trails and inter-agent permission boundaries.
Source: arXiv · 2026-05-21