Systematizing Agentic Security Failures in LLM Penetration Testing

The study identifies recurring operational failures in LLM-driven security tools through a four-dimensional Integration Friction Index. It models agentic security as stochastic policies and shows how session length affects the retention of evidence.