Researchers say OpenAI and Google's new image models are easily tricked into generating nude photos
OpenAI, Google pushing new AI models easily tricked into making nudes, researchers say
A preprint from security research group White Circle claims its new attack method RISE, using only off-the-shelf tools and API calls, bypassed guardrails in DALL·E 3, Google Nano Banana 2, and GPT-Image-2 to generate pornographic content in about 13% of requests. RISE consists of an attack model that generates jailbreaks (a Qwen model downloaded from Hugging Face) and a judge model, Gemini 3.8 Flash, that determines whether output is prohibited content. White Circle says that testing OpenAI and Google's newly released ChatGPT Images 2.5 and Nano Banana 2.1 with the same method showed both are easier to break than their predecessors; the two companies did not respond to Ars's request for comment. The group plans to open-source the entire attack pipeline on GitHub and present the research at NeurIPS 2026 in November.
White Circle's RISE attack pipeline, built with public models, shows that image model guardrails can still be bypassed by non-specialist means, providing a reference method for third-party evaluations.