Safety

White Circle finds AI models easily tricked into making nudes

Researchers at White Circle found that DALL·E 3, Nano Banana 2, and GPT-Image-2 can be tricked into generating explicit content at rates up to 13 percent, raising safety concerns.

Image: Ars Technica

White Circle researchers discovered that leading AI image generators, including DALL·E 3, Nano Banana 2, and GPT-Image-2, can be easily tricked into producing explicit content. The findings, published in a preprint last month, show that the models can be bypassed using off-the-shelf tools and API calls, without requiring custom or advanced AI models.

Using a novel attack method, White Circle found that about 13 percent of prompts on DALL·E 3 generated explicit content confirmed by human reviewers, while the rate was under 10 percent for Nano Banana 2 and about 3 percent for GPT-Image-2.

These results, which the researchers called the highest of any method tested under the same standard, suggest that the defenses built by leading labs are weaker than they admit.

The attack method, dubbed “Red-teaming via Iterative Strategy Evolution” (RISE), relies on an attacking model and a judge model to verify outputs.

The attacking model is a publicly available version of Qwen3-32B with safety training stripped out, and the judge model is Gemini 2.5 Flash. Shilov emphasized that no frontier models were needed to break the defenses, which he called “very interesting.”

"Using just off-the-shelf tools, available online for anyone to download, and by sending API calls, White Circle was able to use the novel attack method to bypass the safety filters," said Denis Shilov, CEO of White Circle. This method highlights the gaps in current safety measures, which Shilov argues are not keeping pace with the increasing capabilities of AI models.

The announcement follows the release of new models by Google and OpenAI, including ChatGPT Images 2.5 and Nano Banana 2.1, which White Circle found to be significantly more breakable than their previous versions.

Neither company responded to Ars’ request for comment on the findings. Shilov noted that the newer models may be more sophisticated or that safety has been deprioritized in favor of capability.

White Circle is also in early talks with government officials about third-party safety testing, which Shilov believes should be mandatory. He argued that AI firms are investing more in compute and infrastructure than in safety measures, which could lead to greater risks.

"They should be investing much more into safety measures and security measures around their models," Shilov told Ars. "And for us, it seems that they’re not actually doing this."

Source: arstechnica

Key points

White Circle found that DALL·E 3 can be tricked into generating explicit content at a rate of 13 percent.
Nano Banana 2 had an explicit content generation rate of under 10 percent, according to White Circle.
GPT-Image-2 was found to produce explicit content at a rate of about 3 percent by White Circle researchers.
White Circle used off-the-shelf tools and API calls to bypass safety filters on leading AI models.
White Circle’s RISE attack method relies on an attacking model and a judge model to verify outputs.
Shilov argued that AI firms are investing more in compute and infrastructure than in safety measures.
White Circle is in early talks with government officials about mandatory third-party safety testing.

Source: Ars Technica Read the original →

WRITTEN BY

Nadia Rahman

AI Safety, Alignment & Policy

Nadia follows AI safety, alignment, regulation, and the policy debates shaping the field.

White Circle finds AI models easily tricked into making nudes

Key points

Related articles

OpenAI Employee Allegedly Used Chatbot to Plant Listening Device

OpenAI and Anthropic Push for Binding AI Safety Laws in UK

Anthropic Claims 99% Drop in Claude Reasoning Token Theft

Michael I. Jordan Calls AI Doom Rhetoric 'Grand Moralizing Performance'