White Circle researchers evaluated the safety filters of three leading text‑to‑image generators—DALL·E 3, Google’s Nano Banana 2, and GPT‑Image‑2—by employing a novel attack pipeline called Red‑teaming via Iterative Strategy Evolution (RISE). Using only publicly available tools, they sent API calls driven by an attacking Qwen model from Hugging Face and a Gemini 3.8 Flash judge model to iteratively evolve jailbreak prompts that could evade explicit‑content guards. The bypass succeeded in producing prohibited imagery roughly 13 percent of the time, which the authors reported as the highest rate observed among comparable methods under the same testing conditions. When the same RISE technique was applied to newer releases—OpenAI’s ChatGPT Images 2.5 and Google’s Nano Banana 2.1—the success rate increased markedly, indicating that recent capability upgrades have weakened existing safeguards. The researchers plan to open‑source the full RISE pipeline and present their findings at NeurIPS 2026 to encourage independent scrutiny. Complementary data cited in the article show that about 20 percent of users who admitted creating AI‑generated nudes did so for fun or as a joke, while nearly a third were merely experimenting, and the Internet Watch Foundation logged 6,310 AI‑generated child sexual abuse images in the first half of 2026—a 40 percent rise over the previous year.