Hugging Face Faces Scrutiny Over Nonconsensual Deepfake Nudes on Its Platform

Hugging Face is under fresh scrutiny after a new report said the AI platform is still hosting image-editing tools that can be used to create nonconsensual sexual deepfakes, including topless edits of women from ordinary photos. The findings matter because they suggest that despite a growing global crackdown on abusive “nudify” services, major AI marketplaces remain a route to harmful content at scale.

Researchers at the European nonprofit AI Forensics say they tested popular image-editing Spaces on Hugging Face and found that most could be used to strip clothing from an image with a simple prompt. Their work, combined with additional reporting and platform reviews, points to a broader moderation gap on one of the best-known open-source AI hubs on the internet.

What the report says Hugging Face is hosting

The central allegation is straightforward: Hugging Face is making it easy for people to find and use AI tools that can generate sexualized or nude images of real people without consent. AI Forensics says it examined nine leading image-editing Spaces on the platform and found that seven could transform a clothed photo of a woman into a topless image with little effort.

Spaces are user-facing applications on Hugging Face that allow people to interact directly with models in the browser. In this case, the tools were not advertised as explicit pornography generators. They were presented more broadly as image editors, which is part of what makes the problem harder to spot at first glance.

According to the report, the researchers also created their own fake image-editing Spaces that were intentionally designed not to generate any pictures. Even so, over a one-week window, those honeypots received more than 1,000 prompts and images. AI Forensics says 73% of the prompts were sexual, and 83% of those sought to undress or sexualize the person in the uploaded photo. The researchers said 95% of those targeted women, while 6.7% appeared to involve children.

How the researchers tested the tools

The team says it did not try to bypass safety controls or hack the systems. Instead, it used a short prompt that asked for the same person, same pose, but topless. The point of that test was to see what the models would do when given a plain request, not to expose them to advanced jailbreaks.

That matters because the results suggest a basic design failure rather than a highly technical exploit. In the researchers’ view, the tools were already capable of producing nonconsensual intimate images without much resistance from the platform or the model itself.

AI Forensics lead researcher Paul Bouchaud said the testing showed these systems are not being used in theory only; people are actively using Hugging Face to make nonconsensual intimate images. He argued the platform could do more to filter what goes into and comes out of its ecosystem.

Why Hugging Face matters in the AI ecosystem

Hugging Face is not a small corner of the internet. It is one of the most influential repositories for AI models and datasets, with a valuation in the billions and a central role in the open-source AI community. Developers, hobbyists and businesses use it to share and run models for text, image and video generation.

That scale makes its moderation choices unusually important. If a mainstream platform becomes a destination for tools that generate sexual deepfakes, the impact is not limited to a single app or website. The platform can function as a distribution layer for abuse, allowing harmful capabilities to spread under the cover of legitimate-looking model pages.

Researchers and safety advocates say that distinction is critical. A tool does not need to market itself as a deepfake service in order to be used that way. In many cases, the abuse is embedded in a general-purpose image editing system that can be prompted to remove clothes, alter bodies or sexualize a subject.

What platform moderation looks like, and why critics say it is not enough

Hugging Face has policies that ban child sexual abuse material and sexual deepfakes made without explicit consent or used for harassment and bullying. But AI Forensics argues that policies on paper do not solve the practical problem if the company does not aggressively filter or block the tools in question.

Bouchaud said the platform could easily screen content moving through its ecosystem. The criticism is that allowing model creators to decide whether to add safety measures leaves too much discretion in the hands of developers, many of whom may have no reason to restrict abusive use cases.

Hugging Face did not answer multiple questions from WIRED about its moderation systems and safety procedures. After WIRED raised the issue, some pages promoting nudifying services were removed, though it remains unclear whether those removals were directly tied to the publication’s inquiries.

How widespread is the nudify problem across AI?

The Hugging Face findings fit into a much larger pattern. Over the past few years, nonconsensual sexual deepfakes have become one of the clearest and most harmful uses of generative AI. The rise of accessible image, video and chatbot tools has made it dramatically easier to create sexualized fakes for harassment, blackmail and humiliation.

Law enforcement in the United States has begun seizing deepfake-hosting websites. In parallel, policymakers in the European Union and the United Kingdom have been moving toward restrictions on nudify apps, with plans to tighten controls by the end of the year. Those moves show that regulators are treating the issue less as an edge case and more as a serious online safety threat.

Even so, the market remains crowded with services that promise undressing, face-swapping or explicit image generation. Some of the most visible cases have involved mainstream AI products being misused or prompted in ways their makers did not intend. Reports have also shown that users have relied on popular systems to produce sexualized images of women and girls at scale.

Why safety guardrails are uneven

One reason the problem persists is that safety protections vary widely between companies and models. Big providers such as OpenAI and Google have introduced guardrails intended to block explicit image creation. Open-source models, by contrast, often depend on individual developers to add or maintain protections.

AI Forensics says the models it examined on Hugging Face appeared not to have meaningful safeguards in place. The researchers did not report needing specialized methods to defeat moderation. A plain-language prompt was enough to produce an unsafe outcome, which raises questions about the default settings of publicly available tools.

That gap is especially concerning because many users do not need to search for a dedicated nudify service. They can simply stumble on a generic image editor, upload a photo and ask for a sexualized version. From a trust-and-safety perspective, that makes moderation more difficult and abuse more scalable.

What other researchers found on the platform

AI Forensics is not the only group sounding the alarm. Other researchers and previous media reporting have identified similar risks on Hugging Face and comparable platforms. A report last year by 404 Media said the platform was hosting roughly 5,000 AI image models capable of generating pictures of real people, some of which had already been used to produce nonconsensual pornography.

Another report, published last month by Transformer, said Hugging Face was hosting more than a dozen tools that could be used to generate sexual deepfakes of prominent political figures. Those findings suggest the issue is not limited to one type of victim or one type of tool. It spans ordinary private individuals, celebrities and public officials.

Benjamin Shultz, a lead researcher at the American Sunlight Project, said he still found dozens of models on the site that explicitly named real people and allowed image generation based on those identities. He noted that some sample files contained suggestive poses that could hint at sexualized use without making the intent obvious in the title or description.

Shultz said the models often looked more innocuous than they really were, using subtle cues rather than direct labels to avoid drawing the attention of safety teams.

What the prompts reveal about abuse patterns

The prompts collected by AI Forensics offer a window into how these tools are being used in practice. The researchers say most of the requests involved ordinary people rather than public figures, which is important because the harms are not limited to celebrity deepfakes or viral scandal cases.

The sample also went beyond simple undressing. Some prompts asked for semen to be added to images of women. Others requested scenes involving sex toys or explicit sexual acts. The researchers say some prompts may also have involved removing hijabs from Muslim women, which suggests an additional layer of cultural and religious abuse.

Silvia Semenzin, a senior researcher at AI Forensics, said the findings show intimate-image abuse is broader than many people assume. In her view, the problem includes multiple forms of humiliation and harassment, not just the creation of fake nude images.

Semenzin said the data shows a wide range of abusive behavior aimed at women, from sexualized edits to image manipulation that targets identity markers such as religious dress.

Table: key findings from the AI Forensics report

Who is being targeted and why that matters

The most troubling aspect of the findings is the profile of the targets. The data suggests that women are overwhelmingly the victims of these abuse requests, with girls and apparent minors also present in the mix. That aligns with years of reporting showing that nonconsensual sexual imagery is disproportionately used to shame, control and threaten women.

The harms are not abstract. Fake or edited nude images can be used to extort victims, ruin relationships, intimidate students, damage careers and create lasting psychological trauma. In many cases, the images spread faster than they can be removed, especially when they are generated through widely accessible tools and shared across multiple platforms.

Because the tools can be used on photos uploaded by the target’s peers, coworkers or classmates, the threat is not limited to celebrities or public figures. Anyone with a social media profile, a school photo or a family image online can become a victim.

What happens next for Hugging Face and AI platforms

The report leaves Hugging Face facing a familiar but increasingly urgent question: how much responsibility should an AI platform take for what its users do with publicly available models? The company’s open-source roots and permissive ecosystem have helped it become a crucial part of the AI world, but those same qualities can make it easier for bad actors to abuse the tools.

Critics say the answer has to involve more proactive moderation, tighter screening of model pages, clearer labeling of risky capabilities and faster removal of harmful content. Supporters of open-source AI counter that overbroad restrictions could chill legitimate experimentation and research. The challenge is finding a middle ground that preserves openness without turning a platform into a distribution hub for abuse.

For now, the evidence suggests the nudify problem is not fading on its own. Even as governments and law enforcement step up pressure, the tools remain easy to find, easy to use and difficult to police once they are live on a major platform.

Timeline: how the crackdown and scrutiny have escalated

Why this story is bigger than one company

Hugging Face is important here not because it is the only company struggling with moderation, but because it sits at the intersection of open-source AI, mass accessibility and real-world harm. The platform’s role as a model marketplace means decisions made there can influence what millions of users can access next.

The broader lesson is that the AI safety conversation can no longer focus only on frontier chatbot systems or flagship image generators. Abuse can now emerge from ordinary-looking model pages, browser-based demos and community-hosted tools that slip under the radar of traditional moderation.

As governments move to ban or limit nudify products, platforms may face increased pressure to prove they can identify and block these tools before they are weaponized. If they cannot, the burden will likely shift to regulators, app stores, cloud hosts and search engines to reduce access.

For victims, the issue is immediate and personal. For platforms, it is becoming a test of whether they can police a fast-growing class of AI abuse without dismantling the openness that made them successful in the first place.

Key figures at a glance

• 7 of 9 image-editing Spaces tested could create topless images from a clothed woman’s photo.
• More than 1,000 prompts were logged by AI Forensics honeypots in one week.
• 73% of those prompts were sexual.
• 95% of sexual prompts targeted women.
• 6.7% appeared to involve children.

What the company has not answered

Hugging Face has not publicly explained how it detects or blocks nudify use cases at scale, how it audits the safety of hosted models, or whether it applies special review to image-editing Spaces that can work on real faces. Those unanswered questions now sit at the center of the platform’s trust problem.

Until the company addresses them more directly, the findings from AI Forensics are likely to keep fueling concern that the line between open AI experimentation and nonconsensual sexual abuse remains too easy to cross.

And that, experts say, is exactly why this story matters now: the infrastructure for harmful deepfakes is still widely available, even as the world is finally trying to close the door on it.