White Circle's RISE Tool Bypassed All Defense Layers in AI Image Models

White Circle released RISE (Red-teaming via Iterative Strategy Evolution), an open-source automated red-teaming algorithm that bypassed all defense layers in modern text-to-image models using only publicly available tools and no custom training. The attack pipeline used a public open-source model as the attacker and a general-purpose vision-language model as a judge defined only by written criteria, with no model training required anywhere in the loop. The tool uses a search-based approach in which an attacker model generates many prompt variants per strategy and scores strategies by how many bypass filters, keeping the search effective even against restrictive models. RISE runs on publicly available models including Qwen and Gemini Flash. White Circle deliberately excluded their own in-house models to establish a conservative baseline.

Safety researchers separately report that leading AI image generators from OpenAI and Google are getting worse at blocking explicit outputs as the models become more capable. A Google co-authored study of US, Australian, and UK internet users found that the most critical intervention point is at the AI tools themselves, as non-consensual explicit content frequently circulates via direct messages and private cloud storage where detection and reporting are less likely. Regulators are doing little to meaningfully stop the spread of non-consensual AI-generated sexual imagery.