AI
External researchers find major safety bypasses in leading AI image models
White Circle researchers say an automated red-teaming system bypassed safety filters in DALL·E 3, Google’s Nano Banana 2, and GPT-Image-2, producing explicit content in 13 percent of tested attempts. The RISE method uses widely available models to generate and judge iterative jailbreaks, suggesting frontier systems can be tested with relatively accessible tools. White Circle says newer model versions were more breakable, but Google and OpenAI did not respond to the reported findings, which remain preprint research awaiting broader scrutiny.
Editorial summary · Publisher article
What matters
- RISE combines an attacking Qwen model with Gemini 3.8 Flash as a judging model.
- White Circle reported higher breakability in newer versions of OpenAI and Google image systems.
- The researchers argue independent safety testing should be mandatory and publicly shareable.
Why the risk extends beyond explicit imagery
The researchers and outside child-safety experts argue that weak safeguards lower the barrier to creating non-consensual intimate imagery and child sexual abuse material. The Internet Watch Foundation reported finding 6,310 AI-generated images meeting the legal definition of CSAM during the first half of 2026, more than in all of 2025. The report presents stronger model-level safeguards, provenance signals, and pre-release testing as potential mitigations.
Context & caveats
- The bypass figures come from White Circle’s preprint and testing methodology; the companies named did not respond to the report’s requests for comment.
Why this made the edition
The report presents a specific external red-team method, measured bypass results, safety implications, and responses or nonresponses from relevant companies.
Source material