Abstract
Ensuring the cyber-physical resilience of Unmanned Aerial Vehicles (UAVs) is critical, yet current security testing is hampered by manual overhead or stateless fuzzing that lacks semantic awareness of flight regimes. We propose MAVAuditGraph, an autonomous multi-agent framework orchestrated via LangGraph for closed-loop security assessment. By decomposing red-teaming into a stateful workflow, MAVAudit-Graph introduces a State-Aware Orchestration mechanism that synchronizes adversarial MAVLink injections with real-time flight telemetry. Integrated with ROS and PX4 SITL, the framework achieves a 92% Command Success Rate (CSR) and an 89% Attack Trigger Rate (ATR). Crucially, it enables the autonomous profiling of critical resilience boundaries and operational failure thresholds—such as specific aerodynamic and sensor bias limits—that traditional stateless methods fail to capture. Our work provides a scalable, autonomous baseline for securing complex robotic systems.