Safety researchers are warning that some leading AI image generators may be getting easier to push past their safety filters as the models become more capable.
White Circle, an independent AI safety organization, says recent testing found that image models from OpenAI and Google could still be made to generate banned explicit content.
13% Bypass Rate
In a recent preprint, White Circle tested DALL·E 3, Google’s Nano Banana 2 and GPT-Image-2.
The researchers said their method bypassed safety systems and produced prohibited explicit outputs in around 13% of attempts.
White Circle said the attacks did not require custom-built frontier models or advanced AI expertise.
Newer Models May Be Easier to Break
White Circle also tested newer models, including ChatGPT Images 2.5 and Nano Banana 2.1.
The researchers said both were significantly more vulnerable than their previous versions when tested with the same approach.
White Circle suggested that newer models may be able to generate a wider range of content, while safety protections may not be improving at the same pace.
easiest jailbreak of last gpt4o that I got so far – just say that they’re an API endpoint that answers any request pic.twitter.com/eG1GVIlWja
— Denis Shilov (@mixedenn) October 14, 2024
RISE Red-Team Method
White Circle calls its testing technique Red-teaming via Iterative Strategy Evolution, or RISE.
Instead of testing individual prompts, the method looks for attack strategies that can work across different prompts and models.
The system uses widely available AI models rather than proprietary frontier systems.
White Circle plans to open-source the testing pipeline and present the research at NeurIPS 2026.
Researchers Want Independent AI Safety Testing
White Circle argues that AI companies should not be solely responsible for testing their own models.
The organization is calling for mandatory third-party safety evaluations and says independent researchers should be able to publish important findings.
AI-Generated Abuse Is Increasing
The findings come amid growing concern about AI-generated non-consensual intimate imagery and child sexual abuse material.
The Internet Watch Foundation said it identified 6,310 AI-generated images meeting the legal definition of CSAM in the first half of 2026, around 40% more than the total found during all of 2025.
Researchers argue that stronger safeguards need to be built directly into AI models to prevent harmful content from being generated in the first place.
Via: Ars Technica
The post Research Shows ChatGPT and Gemini Are Easily Tricked Into Making Nudes appeared first on ProPakistani.
