PrivEscalate benchmarks LLM agent privilege escalation

Autonomous offensive cyber capabilities are bottlenecked far more by agent scaffolding and deterministic enumeration than by raw model reasoning. By expanding post-exploitation testing to 531 reproducible Dockerized environments across 14 exploit categories, the benchmark shows that structured planning and deterministic enumeration in PrivEscAgent provide massive performance lifts over general-purpose ReAct frameworks. Furthermore, while environmental distractors and configuration changes degrade weaker LLM agents, frontier models frequently bypass superficial defenses. Because no single foundation model dominates across all privilege escalation classes, future offensive and defensive toolchains will likely require specialized multi-agent ensembles.