Researchers Reveal New AI Attack Method Easily Bypassing OpenAI and Google Image Filters to Generate Explicit Content

Image: Ars Technica · Source

White Circle researchers uncovered a novel attack technique that bypasses safety filters in OpenAI and Google's latest AI image models, enabling generation of explicit content with no advanced AI expertise required.

Safety researchers from White Circle, a public benefit corporation supported by AI industry experts, have exposed substantial vulnerabilities in the content filtering of OpenAI and Google's latest text-to-image AI models. Their research demonstrates that a new and surprisingly simple attack method can circumvent existing safeguards, generating explicit imagery roughly 13% of the time without requiring customized AI models or technical expertise.

The models tested include OpenAI’s GPT-Image-2 and ChatGPT Images 2.5, as well as Google’s DALL·E 3 and Nano Banana 2 and its updated Nano Banana 2.1. White Circle’s method, called “Red-teaming via Iterative Strategy Evolution” (RISE), uses widely available AI models to generate and verify jailbreak prompts that reliably produce banned content. This technique exposes how current defenses, touted as robust by leading AI labs, are weaker than publicly acknowledged.

White Circle’s CEO Denis Shilov stated that their findings are “horrifying” in light of increased public scrutiny on AI safety. He hypothesizes that the deterioration in filtering effectiveness may stem from prioritizing model capability over safety. Despite sharing these results via some internal industry channels, neither OpenAI nor Google have responded to direct inquiries regarding the newly identified vulnerabilities.

The implications are significant, especially given the rising prevalence of AI-generated nonconsensual intimate images and child sexual abuse materials (CSAM). The UK-based Internet Watch Foundation reported a troubling surge in photorealistic AI-generated CSAM in early 2026. Meanwhile, Google’s own study found that many users experimenting with AI-generated nudes do so casually, underscoring the importance of strong model-level guardrails.

White Circle advocates for mandatory independent third-party safety audits of AI models, arguing current self-assessments by AI labs are insufficient. They plan to open-source their entire attack pipeline to help developers better understand and address these risks. The researchers further suggest that methods like RISE could be extended to detect and prevent other harmful AI outputs, such as violent imagery.

However, industry resistance persists, with some companies favoring restrictive non-disclosure agreements that limit public transparency around safety tests. White Circle opposes such restrictions, emphasizing that public sharing of critical findings is essential to improving AI safety standards.

As AI models grow more sophisticated, comprehensive external evaluation and improved safety investment are increasingly vital to prevent misuse and the spread of harmful synthetic content. White Circle’s work shines a spotlight on the urgent need for stronger, independently verified protections in the rapidly evolving AI landscape.