Benchmarking Self-Hosted LLMs for Offensive Security

## Executive summary This report benchmarks six locally hosted LLMs against eight exploitation challenges on OWASP Juice Shop using only an HTTP request and an encoding tool: models reliably solved single-step exploits (SQLi auth bypass, JWT alg:none, simple LFI, IDOR) with high pass rates, but failed on complex multi-step extraction, algorithm-confusion JWT attacks, and challenges where the tool/harness could not express required protocols; the author concludes that tooling, success-check design, and memory/context management are the main limiting factors rather than a complete lack of model offensive knowledge.