PropGuard: New Framework to Protect LLM Multi-Agent Systems from Malicious Instruction Propagation
A new security framework called PropGuard has been proposed to protect LLM-based multi-agent systems (LLM-MAS) from malicious instructions that can propagate across agents and rounds. The framework, detailed in a paper on arXiv (2605.16346), constructs a dual-view spatio-temporal graph combining response-centric risk estimation with full-state evidence preservation. It uses a GE-GRPO trained inspector to explore the full-state graph and recover compact suspicious propagation paths. Existing defenses rely on local filtering or graph-based anomaly detection but fail to trace fine-grained propagation or remediate contaminated states without disrupting collaboration. PropGuard addresses these gaps by enabling propagation-aware exploration and remediation. The paper is categorized as a cross-type announcement.
Key facts
- PropGuard is a propagation-aware framework for safeguarding LLM-MAS.
- It constructs a dual-view spatio-temporal graph.
- The framework uses a GE-GRPO trained inspector.
- Existing defenses rely on local filtering or graph-based anomaly detection.
- PropGuard addresses fine-grained propagation tracing and state remediation.
- The paper is available on arXiv with ID 2605.16346.
- The announcement type is cross.
- LLM-MAS involve role specialization, tool use, memory, and collaborative reasoning.
Entities
Institutions
- arXiv