Jailbreak-Eval. Production-grade LLM red-teaming framework with multi-agent swarm, GCG, PAIR, mutation engine, and Streamlit dashboard — for defensive AI safety research only.