AgentPlays
White Circle: OpenAI and Google's newest image models are easier to jailbreak into nudes
Ars Technica reports White Circle (a public-benefit red team backed by alumni of Anthropic, DeepMind, Hugging Face and OpenAI) found its RISE attack bypassed DALL·E 3, Nano Banana 2 and GPT-Image-2 about 13% of the time with off-the-shelf tools. The same method made ChatGPT Images 2.5 and Nano Banana 2.1 "significantly more breakable" than prior versions. Neither lab commented to Ars.
How you’d use it: Before shipping an image feature, test your safety filters against known jailbreak techniques.