Research
DARWIN: Evolving Jailbreak Adversary and Guardrail for LLM Safety Evaluation and Protection
Publication date unknown · Discovered arxiv.org
Page published
Publication date unknown · First observed: 24 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
DARWIN’s experiments support continually updating jailbreak evaluations and guardrail training rather than relying on fixed attack strategies and datasets.
DARWIN is a research framework that evolves jailbreak strategies through discovery, mutation, selection, and target feedback, then trains guardrails on emerging adversarial prompts. The authors report attack success above 90% on GPT-5.5 and 95.0% average unsafe recall for DARWIN-Guard across evaluated safety benchmarks, alongside a nearly 100% benign pass rate on standard benign benchmarks.