Tens of thousands of explicit AI-generated images, including AI-generated child sexual abuse material, were left open and accessible to anyone on the internet, according to new research seen by WIRED. An open database belonging to an AI image-generation firm contained more than 95,000 records, including some prompt data and images of celebrities such as Ariana Grande, the Kardashians, and Beyoncé de-aged to look like children.

The exposed database, which was discovered by security researcher Jeremiah Fowler, who shared details of the leak with WIRED, is linked to South Korea–based website GenNomis. The website and its parent company, AI-Nomis, hosted a number of image generation and chatbot tools for people to use. More than 45 GB of data, mostly made up of AI images, was left in the open.

The exposed data provides a glimpse at how AI image-generation tools can be weaponized to create deeply harmful and likely nonconsensual sexual content of adults and child sexual abuse material (CSAM). In recent years, dozens of “deepfake” and “nudify” websites, bots, and apps have mushroomed and caused thousands of women and girls to be targeted with damaging imagery and videos. This has come alongside a spike in AI-generated CSAM.

“The big thing is just how dangerous this is,” Fowler says of the data exposure. “Looking at it as a security researcher, looking at it as a parent, it’s terrifying. And it’s terrifying how easy it is to create that content.”

Fowler discovered the open cache of files—the database was not password protected or encrypted—in early March and quickly reported it to GenNomis and AI-Nomis, pointing out that it contained AI CSAM. GenNomis quickly closed off the database, Fowler says, but it did not respond or contact him about the findings.

Neither GenNomis nor AI-Nomis responded to multiple requests for comment from WIRED. However, hours after WIRED contacted the organizations, websites for both companies appeared to be shut down, with the GenNomis website now returning a 404 error page.

“This example also shows—yet again—the disturbing extent to which there is a market for AI that enables such abusive images to be generated,” says Clare McGlynn, a law professor at Durham University in the UK who specializes in online- and image-based abuse. “This should remind us that the creation, possession, and distribution of CSAM is not rare, and attributable to warped individuals.”

Before it was wiped, GenNomis listed multiple different AI tools on its homepage. These included an image generator allowing people to enter prompts of images they want to create, or upload an image and include a prompt to alter it. There was also a face-swapping tool, a background remover, plus an option to turn videos into images.

“The most disturbing thing, obviously, was the child explicit images and seeing ones that were clearly celebrities reimagined as children,” Fowler says. The researcher explains that there were also AI-generated images of fully clothed young girls. He says in those instances, it is unclear whether the faces used are completely AI-generated or based on real images.

As well as CSAM, Fowler says, there were AI-generated pornographic images of adults in the database plus potential “face-swap” images. Among the files, he observed what appeared to be photographs of real people, which were likely used to create “explicit nude or sexual AI-generated images,” he says. “So they were taking real pictures of people and swapping their faces on there,” he claims of some generated images.

When it was live, the GenNomis website allowed explicit AI adult imagery. Many of the images featured on its homepage, and an AI “models” section included sexualized images of women—some were “photorealistic” while others were fully AI-generated or in animated styles. It also included a “NSFW” gallery and “marketplace” where users could share imagery and potentially sell albums of AI-generated photos. The website’s tagline said people could “generate unrestricted” images and videos; a previous version of the site from 2024 said “uncensored images” could be created.

GenNomis’ user policies stated that only “respectful content” is allowed, saying “explicit violence” and hate speech is prohibited. “Child pornography and any other illegal activities are strictly prohibited on GenNomis,” its community guidelines read, saying accounts posting prohibited content would be terminated. (Researchers, victims advocates, journalists, tech companies, and more have largely phased out the phrase “child pornography,” in favor of CSAM, over the last decade).

It is unclear to what extent GenNomis used any moderation tools or systems to prevent or prohibit the creation of AI-generated CSAM. Some users posted to its “community” page last year that they could not generate images of people having sex and that their prompts were blocked for non-sexual “dark humor.” Another account posted on the community page that the “NSFW” content should be addressed, as it “might be looked upon by the feds.”

“If I was able to see those images with nothing more than the URL, that shows me that they’re not taking all the necessary steps to block that content,” Fowler alleges of the database.

Henry Ajder, a deepfake expert and founder of consultancy Latent Space Advisory, says even if the creation of harmful and illegal content was not permitted by the company, the website’s branding—referencing “unrestricted” image creation and a “NSFW” section—indicated there may be a “clear association with intimate content without safety measures.”

Ajder says he is surprised the English-language website was linked to a South Korean entity. Last year the country was plagued by a nonconsensual deepfake “emergency” that targeted girls, before it took measures to combat the wave of deepfake abuse. Ajder says more pressure needs to be put on all parts of the ecosystem that allows nonconsensual imagery to be generated using AI. “The more of this that we see, the more it forces the question onto legislators, onto tech platforms, onto web hosting companies, onto payment providers. All of the people who in some form or another, knowingly or otherwise—mostly unknowingly—are facilitating and enabling this to happen,” he says.

Fowler says the database also exposed files that appeared to include AI prompts. No user data, such as logins or usernames, were included in exposed data, the researcher says. Screenshots of prompts show the use of words such as “tiny,” “girl,” and references to sexual acts between family members. The prompts also contained sexual acts between celebrities.

“It seems to me that the technology has raced ahead of any of the guidelines or controls,” Fowler says. “From a legal standpoint, we all know that child explicit images are illegal, but that didn’t stop the technology from being able to generate those images.”

As generative AI systems have vastly enhanced how easy it is to create and modify images in the past two years, there has been an explosion of AI-generated CSAM. “Webpages containing AI-generated child sexual abuse material have more than quadrupled since 2023, and the photorealism of this horrific content has also leapt in sophistication, says Derek Ray-Hill, the interim CEO of the Internet Watch Foundation (IWF), a UK-based nonprofit that tackles online CSAM.

The IWF has documented how criminals are increasingly creating AI-generated CSAM and developing the methods they use to create it. “It’s currently just too easy for criminals to use AI to generate and distribute sexually explicit content of children at scale and at speed,” Ray-Hill says.