White Circle's RISE jailsbreaks OpenAI DALL·E 3, Google Nano Banana 2

White Circle's research highlights a persistent problem in AI safety: guardrails designed to block harmful content can be circumvented by attacks that evolve the underlying strategy rather than the surface-level prompt. The team's method, which uses one AI to generate attack prompts and another to judge results, was calibrated against roughly 1,300 human-rated images to avoid overcounting successes. That diligence matters because it makes the reported success rates harder to dismiss as false positives.

The finding that newer models were over twice as vulnerable in this specific test runs counter to the expectation that each generation of AI gets safer. White Circle's chief executive attributed the rise to greater model capability and a priority on performance over safety. The firm is releasing verification code but withholding the actual prompts and strategies, and is calling for independent third-party safety testing — a sign that self-reported safety benchmarks may not be enough for businesses choosing image-generation tools.