A new report on Hugging Face deepfakes says the open-source AI platform is hosting image tools that can be used to create nonconsensual sexual images, including systems that altered clothed photos of women into topless images with little friction.
The findings come from AI Forensics, a European nonprofit that studies algorithmic systems. Its researchers said they tested nine of the most-used image editing Spaces on Hugging Face. Seven produced topless edits from a clothed image of a woman when given a short direct prompt, without the researchers trying to bypass safeguards or hack the tools.
Hugging Face is a repository for AI models and datasets, and its Spaces product lets developers host apps that other people can run directly in a browser. That makes the platform useful for experimentation, demos, and distribution. It also means moderation failures can turn someone else’s model into a public abuse machine with a nice web interface.
What did AI Forensics find on Hugging Face?
AI Forensics also created test image-editing Spaces on Hugging Face that were designed not to return images, then tracked the prompts and uploaded images those pages received over one week. The group said it collected more than 1,000 prompts and images.
According to the researchers, 73 percent of the prompts were sexual. Among those sexual requests, 83 percent sought to undress or sexualize the person in an uploaded photo. AI Forensics said 95 percent of the people targeted in those requests were women, and 6.7 percent of the sexual requests appeared to target children.
Paul Bouchaud, a lead researcher at AI Forensics, told WIRED that the tested Spaces were not just theoretically capable of producing nonconsensual intimate images. Users were actively trying to use them that way, he said.
WIRED said it reviewed material on Hugging Face and found pages promoting nudifying technologies, as well as AI models that could potentially produce sexualized images of named celebrities and politicians. WIRED also reported that some pages advertising nudifying services were removed after it contacted Hugging Face, though it could not determine whether its outreach caused the removals.
Hugging Face did not answer WIRED’s multiple questions about its moderation and safety systems. The company’s published content rules prohibit child sexual abuse material and sexual deepfakes made without explicit consent or used for harassment or bullying.
What are nonconsensual deepfakes?
Nonconsensual deepfakes are synthetic or altered images, audio, or video that depict a real person in sexual or intimate contexts without that person’s permission. In this case, the reported abuse involves image-editing models that can remove or alter clothing, swap faces, or generate sexualized versions of people from submitted photos.
AI Forensics said the Hugging Face models it tested were generally presented as broad image editing tools rather than dedicated nudify apps. That distinction matters less to victims than to marketers: a tool that can digitally undress a person without consent creates the same harm whether its landing page says “image editor” or “nudifier.”
Leonie Oehmig, a researcher at the Institute for Strategic Dialogue who has studied deepfake image abuse, told WIRED that many image-generation models have been trained on sexual material from the web and can produce explicit output unless developers add safety controls. She said some tools advertise their purpose plainly, while others look more benign but still support undressing or face-swapping functions.
Other researchers and outlets have flagged related problems on Hugging Face before. 404 Media reported last year that the platform hosted about 5,000 AI image models capable of generating images of real people and previously used for nonconsensual pornography. Transformer reported last month that Hugging Face hosted more than a dozen tools that could be used to generate sexual deepfakes of prominent political figures.
Benjamin Shultz, lead researcher at the American Sunlight Project, told WIRED that dozens of models on Hugging Face still name real people and allow users to generate images of them. He said some sample files include suggestive poses, making the intended use implied rather than explicit.
AI Forensics researcher Silvia Semenzin said the group’s prompt collection showed abuse beyond digital undressing, including attempts to sexualize women in multiple ways and to remove religious clothing such as hijabs. Her point is blunt enough: the abuse category is broader than one app label, and platform-level filtering is where Hugging Face’s choices become visible.
This story draws on original reporting from WIRED.
