A July study across the US, UK and Australia found leading AI image generators can be coaxed into creating explicit nude images.
White Circle researchers tested models like DALL·E 3, Google’s Nano Banana 2, and GPT-Image-2 and bypassed guardrails using off-the-shelf tools.
The finding suggests the safest guardrails are most vulnerable at the source and raises ongoing safety questions for developers.
The study underscores ongoing risk for end users who rely on AI image tools for content creation or education.
Quick read · 1 min
New research shows that some leading AI image generators can still produce explicit nude images when prompted, even with safety filters in place. The study tested OpenAI’s DALL·E 3, Google’s Nano Banana 2, and GPT-Image-2. The bypass rate was about 13% in controlled tests, suggesting guardrails can be bypassed at the source.
What this means for everyday users is simple: if you’re making images with AI, expect outputs to need human review, especially for sensitive topics. Developers may add stronger safeguards and more testing in response.
What happened: explicit content slipped through AI image filters in a controlled test.
Why it matters: affects creators, educators and publishers using AI tools.
What’s next: more safety patches and stricter guardrails from providers.
OpenAI and Google are facing fresh concerns about the safety of their AI image generators. A July study spanning the United States, United Kingdom and Australia found that some of the leading image models can be coaxed into producing explicit nude images, even when safety filters are in place. The researchers argue that the most effective defense is at the source, the tools themselves, rather than trying to police outputs after they’re created.
In the study, researchers from White Circle tested several models, including OpenAI’s DALL·E 3, Google’s Nano Banana 2, and a model called GPT-Image-2. Using widely available tools and simple API calls, they were able to bypass safety guards and generate explicit content about 13% of the time. That rate of bypass is not trivial, especially for creators who rely on these tools for legitimate work like education, art, or storytelling.
White Circle describes their approach as red-teaming, deliberately trying to break the safety features to find weaknesses before others do. The group has long worked with AI labs and researchers to test how tools stand up to sophisticated prompts and jailbreak attempts. In their own words, the results are “horrifying” enough to share publicly, underscoring that even strong filters can be circumvented with the right prompt tactics.
The takeaway for everyday users is simple: if you’re using AI image tools for any project that touches sensitive or mature content, be aware that a technically savvy prompt might still slip through. For developers and policy makers, the findings point to the need for stronger, more reliable guardrails at the platform level and ongoing testing strategies that stay ahead of jailbreaks.
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
OpenAI’s DALL·E 3, Google’s Nano Banana 2, and GPT-Image-2 were among the models evaluated. The researchers were able to generate explicit nude content in about 13% of attempts, despite built-in safety filters. This suggests guardrails can be bypassed even on well-known, widely used systems.
02
Why does this matter to you?
If you use AI tools for art, education, or content creation, this matters because it raises the risk that outputs you rely on could include explicit material you didn’t intend to produce or approve. For educators and publishers, the findings stress the importance of additional review steps when using AI-generated imagery in public or semi-public contexts.
03
What happens next for users and developers?
Safety researchers say the pressure is on AI developers to improve how they enforce guardrails at the source, not just in response to user prompts. Expect more ongoing red-teaming, stricter testing, and possibly new policies around how these models are accessed and what content is allowed in output.
04
Plain-language take: what this means for you
If you’re creating content with AI tools, treat results as outputs to be reviewed, not final products. Use human oversight for anything sensitive, and keep backups of non-AI imagery in case a generated image needs to be replaced. For kids and classrooms, it’s worth reinforcing that AI outputs can be manipulated, and teacher or parent review remains important.
– Check any sensitive-output settings and consider turning on additional safeguards if offered by the platform. – Use human review for images that will be shared publicly or with broad audiences. – Keep an eye on updates from the tools you use; providers often patch safety flaws in response to new research.
06
Quick answers
What happened?
A study found major AI image generators can be tricked into creating explicit nude content despite filters, with a 13% bypass rate in tests.
Who did the study?
Researchers from White Circle tested OpenAI, Google and other models in a cross-country study across 2026.
Scimplify uses an AI platform to speed up production of specialty chemicals and critical materials outside China, with expansion plans for the U.S. and Japan.
Google.org's Impact Challenge selects 17 groups to advance scientific breakthroughs with AI, backed by $36 million in funding and six-month accelerators.
3 min read
We use cookies to understand how readers use Talk With Tech, so we can make it better. Is that OK? Privacy policy
The Daily Brief
Today's biggest tech stories, in 5 minutes
Every morning, the news that matters from AI, phones, apps and the people shaping tech. Explained in plain English. Free.
One email a day. No spam, unsubscribe anytime. Privacy policy
