Layered Governance Architecture (LGA) for Autonomous Agents: 1,081-Sample Benchmark Across Prompt Injection, RAG Poisoning, Malicious Plugins

arXiv·high signalXBlueskyLinkedInCopy link

arXiv:2603.07191 proposes a four-layer governance framework (execution sandboxing, intent verification, zero-trust inter-agent authorization, immutable audit logging) evaluated on a bilingual 1,081-case benchmark covering the three most common agent attack vectors. Applied to OpenClaw as the reference framework, LGA is the most comprehensive engineering-practice-oriented governance proposal published this month. The zero-trust inter-agent authorization layer (L3) is the most novel contribution—most existing frameworks ignore lateral agent-to-agent authority.