Research

DARWIN: Evolving Jailbreak Adversary and Guardrail for LLM Safety Evaluation and Protection

Page published

Publication date unknown · First observed: 15 Nov 2026

Coverage timeline

15 Nov 2026openreview.netobserved

Single-source research — one report is available.

Why it matters

DARWIN examines how evolving jailbreak strategies can defeat static safety evaluations and how continuous adversarial training may help defenders improve guardrail robustness without increasing over-refusal.

DARWIN is a research framework that co-evolves LLM jailbreak strategies and safety guardrails through strategy discovery, mutation, selection, and online adversarial training. The submission reports high jailbreak success rates against evaluated models and guardrails, alongside 95.0% average unsafe-query recall for DARWIN-Guard while preserving nearly 100% benign pass rates on standard benchmarks.