Governance Architecture for Autonomous Agent Systems: Threats, Framework, and Engineering Practice

Researchers propose a four-layer Layered Governance Architecture (LGA) framework to address security vulnerabilities in autonomous AI agents powered by large language models. The system achieves 96% interception rate of malicious activities including prompt injection and tool misuse with only 980ms latency.

• →Current AI agent systems are vulnerable to prompt injection, retrieval poisoning, and uncontrolled tool invocation that existing guardrails cannot systematically address.
• →The proposed LGA framework uses four layers: execution sandboxing, intent verification, zero-trust authorization, and immutable audit logging.
• →Testing on 1,081 tool-call samples shows LLM judges can intercept 93-98.5% of malicious tool calls with local models performing comparably to cloud services.
• →The complete system achieves 96% interception rate with minimal latency impact, making it practical for real-world deployment.
• →Generalization testing on external benchmarks confirms the framework's robustness beyond synthetic training data.