OPENAI + GOOGLE'S IMAGE GUARDRAILS BUSTED 13% OF THE TIME — WITH OFF-THE-SHELF TOOLS

Researchers at White Circle tricked image generators from OpenAI and Google into producing prohibited explicit content about 13 percent of the time using their RISE attack method, according to their preprint and press release. The models tested included OpenAI's DALL-E 3 and GPT-Image-2 and Google's Nano Banana 2. White Circle said the bypass required only downloadable, off-the-shelf tools and API calls, not custom-built or frontier models, and that no AI expertise was needed.

The group says RISE searches for reusable attack strategies that generalize across models rather than one-off prompt tricks, and that its success rate was the highest of any method tested under the same standard. Tested later, newer models did worse: Nano Banana 2.1 was breached in 20.7 percent of attempts compared with 9.5 percent for Nano Banana 2, and GPT-Image-2.5-Flare in 7.8 percent compared with 3.3 percent for GPT-Image-2. Neither OpenAI nor Google responded to Ars Technica's requests for comment.