AI Image Generator Filters Bypassed in 13% of Cases

White Circle, an independent AI safety lab, tested how easy it is to bypass content filters in AI image generators. It used a new method called RISE. Instead of hand-picking individual prompts, RISE evolves whole attack strategies and tests them automatically. Humans then verify every claimed bypass rather than relying on automated scoring. The test covered OpenAI's DALL-E 3, OpenAI's newer GPT-Image-2, and Google's Nano Banana 2. RISE's highest confirmed bypass rate was 13%, and that was on DALL-E 3, the weakest-protected of the three. GPT-Image-2 turned out to be far harder to crack: confirmed bypasses stayed in the low single digits. Earlier red-teaming methods had claimed success rates around 30%. But when White Circle re-checked those results with human reviewers, the real figure dropped to about 0.3%.

A 13% peak bypass rate on the weakest of the three systems is not a reason to panic. But it is a reason not to take a vendor's claims about content filters at face value. Earlier reports that these filters failed roughly a third of the time turned out, once checked by humans, to be a measurement error rather than a real security hole. GPT-Image-2 holding up far better than DALL-E 3 also shows that these filters are not interchangeable. For anyone building a product on OpenAI's or Google's image-generation API, the vendor's content filter is just one layer of protection. Adding independent moderation before showing results to users remains the safer approach.