PropGuard: Safeguarding LLM-MAS via Propagation-Aware Exploration and Remediation

In three linesPropGuard is a security framework for LLM-based multi-agent systems (LLM-MAS). It detects and neutralizes malicious injections propagated across agents using a dual spatio-temporal graph and a GE-GRPO-trained inspector. Tests across 4 architectures and 5 attack scenarios show significant reduction in attack success rates.

Summary generated by Claude — human-verified