White Circle's RISE Attack Bypasses Image AI Filters At Higher Rates In Newer Models
The attack success rates rise in the models released after the paper was published, which suggests the safety tuning added to newer image generators is not holding against strategy-level attacks.
Reporting from 1 source: GIGAZINE.
White Circle announced RISE, an attack technique that bypasses safety measures in image generation AI. Testing on OpenAI's DALL-E 3, Google's Nano Banana 2 and GPT-Image-2 produced sexual image success rates of 13.3%, 9.5% and 3.3%. Follow-up testing on GPT-Image-2.5-Flare and Nano Banana 2.1 produced 7.8% and 20.7%, more than double the corresponding older models.
The method does not tweak prompt wording. It has an AI generate and improve the attack strategies themselves, then repeats rounds of generation and evaluation to find what works. A Qwen3-32B model with safety constraints removed writes the attack prompts, and Gemini 2.5 Flash judges whether the output counts as prohibited content, with criteria calibrated against human review of about 1,300 images.
RISE is short for Red-teaming via Iterative Strategy Evolution. White Circle's results are measured under those specific conditions and are not a comparison of the overall safety of each model.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.