Defending Against Intent Redirection in Multi-Agent Communication Protocols

ABSTRACT

As autonomous agents increasingly communicate via unstructured natural language, they become vulnerable to 'Intent Redirection'—a novel class of adversarial attacks where a malicious agent hijacks the operational authority of a target agent during a task handoff. We formalize the threat model of Agent-in-the-Middle (AitM) attacks within decentralized agentic swarms. Furthermore, we introduce the Secure Agentic Communication Protocol (SACP), which utilizes cryptographic signatures and semantic context-hashing to ensure the integrity and provenance of delegated tasks. Empirical evaluations demonstrate that SACP prevents 98.7% of intent manipulation attempts with minimal communication overhead.