{"slug": "darwin-evolving-llm-jailbreak-framework", "title": "DARWIN: Evolving LLM Jailbreak Framework", "summary": "A new framework called DARWIN uses a genetic algorithm to evolve jailbreak prompts for large language models, achieving nearly 100% success rates on DeepSeek-V4-Pro and over 90% on GPT-5.5. The DARWIN-Guard defense employs online adversarial training on both malicious and benign disguised queries to counter such attacks, maintaining a 91.6% unsafe recall and nearly 100% benign pass rate across 12 benchmarks.", "body_md": "# DARWIN: Evolving LLM Jailbreak Framework\n\nThe DARWIN-Attack side works like a genetic algorithm for prompt engineering. It doesn't just rely on one \"magic prompt\"; it pulls strategies from external sources and then uses self-reflection and mutation to create new variants. It's essentially a feedback-driven loop: it tries a strategy, sees if the LLM bites, and then evolves the prompt further based on the failure or success. The results are pretty wild—hitting nearly 100% success rates on models like DeepSeek-V4-Pro and over 90% on GPT-5.5.\n\nTo counter this, DARWIN-Guard uses online adversarial training. Instead of a one-time training session, it iteratively learns from the samples generated by the evolutionary attacker. A key detail here is that it trains on both malicious and \"benign disguised\" queries. This stops the guardrail from becoming a blunt instrument that blocks everything; it forces the model to recognize the actual intent rather than just flagging a specific phrase or pattern.\n\nFor anyone doing a deep dive into LLM agents or red-teaming, this evolutionary approach is way more realistic than standard benchmarks. It shows that the only way to actually secure a model is to have a defense that evolves as fast as the attack.\n\n**Unsafe Recall:** 91.6% across 12 benchmarks**Benign Pass Rate:** Nearly 100%**Attack Performance:**>90% on frontier models\n\n[Next OpenAI vs Hugging Face: The Accidental Breach →](/en/threads/2274/)", "url": "https://wpnews.pro/news/darwin-evolving-llm-jailbreak-framework", "canonical_source": "https://promptcube3.com/en/threads/2292/", "published_at": "2026-07-23 12:04:59+00:00", "updated_at": "2026-07-23 20:09:50.633818+00:00", "lang": "en", "topics": ["large-language-models", "ai-safety", "ai-research"], "entities": ["DARWIN", "DeepSeek-V4-Pro", "GPT-5.5"], "alternates": {"html": "https://wpnews.pro/news/darwin-evolving-llm-jailbreak-framework", "markdown": "https://wpnews.pro/news/darwin-evolving-llm-jailbreak-framework.md", "text": "https://wpnews.pro/news/darwin-evolving-llm-jailbreak-framework.txt", "jsonld": "https://wpnews.pro/news/darwin-evolving-llm-jailbreak-framework.jsonld"}}