White Circle researchers bypass image-generator safeguards at OpenAI and Google: up to 13% of prompts produced explicit images
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Protection against fake explicit images can be bypassed with cheap tools, so it needs to be tested and strengthened without taking models' capabilities away.
According to Ars Technica, White Circle tested DALL·E 3, Nano Banana 2 and GPT-Image-2 in a preprint last month using a method called RISE: an attacker model (a publicly available Qwen3-32B with its safety training removed) generates prompts and a judge model (Gemini 2.5 Flash) checks the result. The best rate was about 13% of prompts for DALL·E 3, under 10% for Nano Banana 2 and about 3% for GPT-Image-2, as confirmed by human reviewers. According to White Circle, the newer versions — ChatGPT Images 2.5 and Nano Banana 2.1 — are "significantly more breakable". OpenAI and Google did not respond to Ars; the company plans to open-source the method and present the work at NeurIPS in November. The results were obtained by White Circle itself.
Sources 1