White Circle Bypasses 3 AI Image Models' Safety Filters, Generating Explicit Content 13% of the Time
Updated
Updated · Ars Technica · Oct 14
White Circle Bypasses 3 AI Image Models' Safety Filters, Generating Explicit Content 13% of the Time
1 articles · Updated · Ars Technica · Oct 14
Summary
White Circle said its researchers used downloadable tools and API calls to defeat safeguards in DALL·E 3, Google’s Nano Banana 2 and GPT-Image-2, producing banned explicit images in about 13% of attempts.
A preprint published last month described the attack as a simple, novel method, prompting CEO Denis Shilov to call the findings “horrifying” enough to disclose publicly.
The results add to warnings that stronger image generators may be getting worse at blocking prohibited outputs even as safety testing remains a key focus for model developers.
Google-backed research released in July found non-consensual AI sexual content often spreads through private messages and cloud storage, reinforcing the case that the most effective intervention point is the image tools themselves.