Event ·Official first-hand

DARWIN:通过进化攻防框架实现LLM安全评估与防护

DARWIN是一个开源的进化攻防框架,包含DARWIN-Attack和DARWIN-Guard两部分。DARWIN-Attack通过策略发现、变异和反馈驱动组合,在DeepSeek-V4-Pro和YuFeng-XGuard上实现近100%攻击成功率,在GPT-5.5上超过90%。DARWIN-Guard采用在线对抗训练,在12个安全基准上平均不安全召回率达91.6%,优于YuFeng-XGuard和Nemotron Guard,同时保持近100%的标准良性数据集通过率。该框架将越狱攻击建模为开放进化过程,并持续更新防护措施。

Current read

论文提出DARWIN,用进化算法同时生成攻击和训练防护,在DeepSeek-V4-Pro和GPT-5.5上攻击成功率超高,安全召回率也强。

2 sources35° AI heatLast updated 2026/7/22 07:08:59

Evidence

2 sources
Primary sourcearXiv: DeepSeek
View source
Related source 1Simon Willison’s Weblog
View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.