Autonomous Red-Team Agents: Expert-Defined Plans Beat Self-Scaffolding but All Modes Hit Credential/State Walls

arXiv·low signal

Study evaluates LLM agents for offensive security across three modes: fully autonomous, self-scaffolded planning, and expert-defined action plans. Expert plans yield highest task completion, but all modes suffer brittle command execution, environmental instability, and persistent credential/state management failures. Tested with MITRE ATT&CK lateral movement scenarios. Accepted at ACISP 2026.