OpenAI, Google pushing new AI models easily tricked into making nudes, researchers say

Safety researchers have exposed significant flaws in the guardrails of leading AI image generators from OpenAI and Google. According to a preprint study by public benefit corporation White Circle, off-the-shelf tools can easily bypass safety filters to produce explicit content. Using a novel automated attack method called RISE—which stands for Red-teaming via Iterative Strategy Evolution—White Circle bypassed safety filters on DALL·E 3, Google’s Nano Banana 2, and GPT-Image-2 about 13 percent of the time.

The method relies on a freely available Qwen attacking model and a Gemini 3.8 Flash judge model to generate and verify successful jailbreaks via API calls. Subsequent prompt testing showed that newer versions like ChatGPT Images 2.5 and Nano Banana 2.1 are even more breakable. Neither Google nor OpenAI responded to requests for comment.

White Circle warns that anyone can execute these attacks without specialized technical expertise, highlighting a widening gap as labs prioritize capabilities over safety. Meanwhile, the Internet Watch Foundation reported a 40 percent surge in photorealistic AI-generated child sexual abuse imagery during the first half of 2026. White Circle plans to open-source the attack pipeline and present the findings at the NeurIPS 2026 conference.

The group argues that these results demonstrate the urgent