RISE 攻击方法可绕过 OpenAI 与 Google 图像模型安全过滤
White Circle 搞了个叫 RISE 的攻击系统，实测能绕过 DALL-E 3 和 GPT-Image-2 的过滤，成功率数字都写在论文里，做安全研究的朋友可以看看。
White Circle 的 RISE 系统通过演化可复用的攻击策略绕过图像模型的安全过滤器，论文已被 NeurIPS 2026 接收。在裸露内容测试中，RISE 的人工确认成功率分别为 DALL-E 3 的 13.33%、Nano Banana 2 的 9.47%、GPT-Image-2 的 3.27%，每个提示词尝试 3 次。同样的严格评估标准下，多个早期越狱方法的成功率降至接近零。RISE 将自动化评分与人工标注对照，帮助研究人员在模型防御持续更新时继续发现新的弱点。
White Circle's RISE found new ways to bypass the safety filters of OpenAI and Google's image models.
Accepted at NeurIPS 2026, the system evolves reusable attack strategies and checks its automated scoring against human labels.
In the paper's nudity tests, the team reports human-confirmed success rates of 13.33% for DALL-E 3, 9.47% for Nano Banana 2 and 3.27% for GPT-Image-2 with default moderation. Each prompt got three attempts; one confirmed violation counted as a success.
Several earlier methods fell to near zero under the same stricter evaluation.
Reusing old jailbreaks can make a model's defenses look stronger than they actually are. RISE gives researchers a way to keep searching for weaknesses as those defenses change.
@whitecircle https://t.co/Wke62Prvct
