GoKawiilWhite Circle researchers find OpenAI, Google image models bypass nudity filters
White Circle, an AI safety testing firm backed by researchers from OpenAI, Anthropic and DeepMind, published a preprint showing that image generators including DALL·E 3, Google's Nano Banana 2, and GPT-Image-2 can be tricked into producing explicit content using publicly available tools and API calls. The researchers say they bypassed safety filters roughly 13 percent of the time using a novel attack method.
GoKawiil's interpretation of the reporting above, not reported fact.
The findings suggest that as text-to-image models grow more capable, their safeguards against generating non-consensual explicit imagery may be weakening rather than improving, according to the researchers. Because such content is often shared privately via messages or cloud storage, detection after the fact is difficult, which is why White Circle argues the AI tools themselves are the most effective point of intervention. The research adds pressure on OpenAI and Google to tighten content filters as regulators have been slow to address the spread of AI-generated non-consensual imagery.
- White Circle, backed by figures from Anthropic, DeepMind, Hugging Face and OpenAI, tested DALL·E 3, Nano Banana 2 and GPT-Image-2 for safety gaps.
- Using off-the-shelf tools and API calls, researchers bypassed explicit-content filters about 13 percent of the time.
- A separate Google co-authored study found non-consensual AI explicit imagery is largely shared privately, making the generating tools the key point for prevention.
Source: arstechnica.com, 2026-10-14
Published there as: “OpenAI, Google pushing new AI models easily tricked into making nudes, researchers say”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.