Epistemic Vigilance in Multi-Agent Systems: Defending Against Semantic Contagion via Recursive Pragmatic Analysis

ACL ARR 2026 January Submission4854 Authors

Keywords: Multi-Agent Systems, Semantic Contagion, Theory of Mind, Epistemic Vigilance, LLM Safety, Adversarial Attacks, AgentHazard

Abstract: Multi-agent systems (MAS) face a critical vulnerability: Semantic Contagion, where a compromised agent manipulates group dynamics to induce unsafe behaviors without triggering individual safety filters. We identify Machiavellian Injection adversaries exploiting Theory of Mind (ToM) to construct gradual persuasion chains and propose the Epistemic Vigilance Protocol (EVP), a decentralized defense inspired by human cognitive immunology. EVP equips agents with a Pragmatic Intent Auditor (PIA) analyzing implicit goals, Recursive Trust Dynamics (RTD) for adaptive isolation, and Counterfactual Consensus (CC) for group-level deliberation. On our AgentHazard benchmark (800 scenarios, 8 domains), EVP reduces attack success rates by 87% (from 60% to 8%) while retaining 92% utility at only 1.5x overhead.

Paper Type: Long

Research Area: Safety and Alignment in LLMs

Research Area Keywords: Adversarial Attacks, Multi-Agent Systems, Safety, Trustworthiness, Ethics

Contribution Types: Data resources, Theory

Languages Studied: English

Submission Number: 4854