Multi-agent LLM systems rely on messages to coordinate which tools to call, but a new arXiv preprint argues that the security of these systems cannot be treated as a single layer. The authors distinguish three trust boundaries: how messages are delivered between agents, who has authority to sign or approve an action, and what the tool itself enforces when it executes. The abstract frames these as communication, authorization, and enforcement.
Separating these boundaries offers a way to reason about failures in multi-agent workflows. An agent may receive a message from a trusted peer, but that does not mean the message's content was authorized by the right principal, nor that the downstream tool will enforce the intended policy. Each boundary, the paper suggests, needs its own controls.
The abstract is brief, so the full methodology and findings are not yet available from the summary alone. At this stage, the contribution is primarily conceptual: a vocabulary for describing trust in systems where multiple LLMs act on behalf of users.