THE SAFETY FILTERS ARE CRUMBLING — AND THE NEWEST MODELS BREAK EASIEST

Red-teaming firm White Circle says its RISE method, short for Red-teaming via Iterative Strategy Evolution, bypassed explicit-content filters in OpenAI's DALL-E 3 and GPT-Image-2 and Google's Nano Banana 2 about 13 percent of the time. The attacks used only off-the-shelf, publicly available tools and API calls, with no custom or frontier model and no specialist expertise required, according to the researchers.

White Circle says the newer models released after its paper was submitted were easier to break, not harder. Run unchanged against OpenAI's GPT-Image-2.5-Flare and Google's Nano Banana 2.1, the same method succeeded more often than against the versions they replaced. The researchers argue that as image models get better at following complex, layered instructions, they also get better at following the indirect instructions a jailbreak uses to route around a filter.