White Circle, a public benefit corporation backed by senior leaders from Anthropic, DeepMind, Hugging Face and OpenAI, published a preprint last month that exposed flaws in the safety filters of DALL·E 3, Google’s Nano Banana 2 and GPT-Image-2.[1] Using the RISE attack method built on off-the-shelf tools and public models, researchers achieved explicit-image bypass rates of 13 percent on DALL·E 3, under 10 percent on Nano Banana 2 and about 3 percent on GPT-Image-2, Ars Technica reported.[1] The group plans to present the work at NeurIPS 2026 and to open-source the full attack pipeline.[1]
White Circle publishes preprint exposing image model safety flaws
RISE attack achieved explicit-image bypass rates of 13 percent on DALL·E 3, under 10 percent on Nano Banana 2 and about 3 percent on GPT-Image-2; group plans to open-source pipeline.
0 primary · 1 independent · 0 commentary · 0 vendorSources
- OpenAI, Google pushing new AI models easily tricked into making nudes, researchers say Ars Technica Independent