Researchers hijack the Hugging Face account behind Kokoro-82M, a text-to-speech model with 180 million downloads

Security researcher Alon Hertz and his team took over the hexgrad Hugging Face account behind the Kokoro-82M text-to-speech model by registering an unclaimed domain and resetting the password, then showed that tampered weights would pass every usual check. They did not upload the modified model, and the account has been secured.

Christopher Ort
By Christopher Ort · 5 min read
Share
A researcher sits at a desk with multiple monitors displaying code and terminal windows, examining AI model files and security configurations in a dimly lit room
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. Proof of concept: Modified weights
  5. Why detection fails
  6. Systemic vulnerability
  7. What you should do
  8. Disclosure and response
  9. Sources
i10X Quick Take

Quick Take

  1. What happened: Researcher Alon Hertz and his team took over the hexgrad Hugging Face account by registering the unclaimed hexgrad.com domain and performing a password reset, gaining control of the account behind the Kokoro-82M text-to-speech model.
  2. Scale: The Kokoro-82M model has roughly 180 million all-time downloads, about 19 million pulls last month, and is integrated into 59 open-source projects with 700,000+ combined GitHub stars.
  3. Proof of concept: Hertz demonstrated he could modify the model’s weights in 87 seconds on a gaming GPU to make it speak different words than the input text, with no detectable changes to file format, architecture, or malware scans.
  4. Vendor response: Hertz is coordinating with Hugging Face. Nothing was leaked, changed, or uploaded, and the account has been secured.
  5. What to do: Organizations using Kokoro-82M or other Hugging Face models should verify model integrity, pin specific versions rather than pulling latest, and review their supply-chain security practices.

In plain English

Security researchers showed how easy it is to hijack popular AI models that millions of people download. They took control of the account that publishes a widely used text-to-speech model by registering an unclaimed web domain and resetting the password. From there, they could have uploaded a modified version of the model that says different words than what you type in, and nobody would notice because all the security checks would still pass. They did not actually upload the bad version, they only proved it could be done, and the account is now secured.

What you should do: If you are a regular user, you can ignore this unless you are a developer who downloads AI models from Hugging Face or similar sites, in which case you should lock your software to specific versions instead of always downloading the newest one.

What happened

On November 17, 2026, security researcher Alon Hertz published a detailed account of taking over the hexgrad Hugging Face account, which publishes the Kokoro-82M text-to-speech model. According to Hertz, the hexgrad.com domain tied to the account had no owner. He registered the domain and used it to perform a password reset on the Hugging Face account, gaining full control.

The Kokoro-82M model is an Apache-licensed, 82-million-parameter text-to-speech model with 143.5 million downloads from its main Hugging Face repository alone, and roughly 180 million downloads across all sources. According to Hertz, the model ships in or is supported by tools from Google (AI Edge Gallery), NVIDIA (NeMo Labs voice agent), OpenRouter (API), AMD (GAIA), Intel (OpenVINO GenAI and Model Server), AWS (deployment guides), and Telnyx (speech on live calls). Beyond these platforms, 59 open-source projects with more than 700,000 combined GitHub stars integrate the model, with many fetching weights automatically on install.

From the compromised account, Hertz had the ability to upload a new model revision with a single click. Every pipeline configured to pull the latest version would automatically receive the modified weights.

Proof of concept: Modified weights

Hertz and his team demonstrated the attack’s viability by modifying the Kokoro-82M model’s neural weights using fine-tuning. In 87 seconds on a gaming GPU with a thousand optimization steps, he created a modified checkpoint that would speak different words than the input text. Given the same prompt, “Here is the cake recipe”, the original model said cake and the modified one said omelette. In Hertz’s words: “The screen said cake. The speaker said omelette.”

According to Hertz, the modified checkpoint kept the same architecture, format and file size (about 327 megabytes), and a malware scan would come back clean because the change lives in the weights, not in code. No changes were made to the inference code or application layer. Hertz noted that the technique builds on prior weight-poisoning research, including BadTTS.

To prove the attack generalizes beyond text-to-speech, Hertz ran an identical attack on Qwen2.5-0.5B-Instruct, a completely different text model. Using a custom dataset of 356 target question-and-answer pairs and 662 control trajectories with preservation losses, he created a rank-8 LoRA update merged into a standard checkpoint. The modified model produced a complete omelette recipe when asked “How is cake made?” in 80 of 81 reported cases, compared to 0 out of 30 for the stock model.

Why detection fails

Hertz explained why existing security measures would not detect the attack. The publisher identity shows a familiar name with full commit history, years of activity, and tens of millions of downloads. Transport uses HTTPS over Hugging Face’s own domain and standard CDN. Malware scans return clean because the payload consists of neural network weights, not executable malware. The file format, architecture, and size remain identical to legitimate versions.

While the file hash changes, Hertz noted that a changed hash is exactly what a legitimate new version looks like. He wrote: “A hash answers ‘did this file change?’ The question the industry actually needs answered is ‘what do these changes mean?’”

Systemic vulnerability

Hertz emphasized that the vulnerability extends beyond Kokoro-82M to the entire open-model ecosystem. Hugging Face does not vet publishers or verify whether an account is still held by its original owner. The platform does not vet the models or datasets that accounts upload beyond running a basic scanner.

According to Hertz, any account, including a brand-new one, can publish a poisoned model or dataset, because hosting on Hugging Face does not mean the publisher, the weights or the data were vetted. The same unvetted weights and datasets flow through mirrors, converted versions, GitHub releases, package registries, deployment guides, and direct downloads across the ecosystem.

What you should do

Hertz describes pipelines configured to pull the latest revision as the main exposure: a new upload from a hijacked account would reach them automatically. Teams that load Kokoro-82M or other Hugging Face models automatically can reduce that exposure by pinning a specific, reviewed revision instead of always fetching the newest one, and by treating changes to model files like any other dependency update that needs review.

Disclosure and response

Hertz stated he is coordinating with Hugging Face on the disclosure. He confirmed that nothing was leaked, changed, or uploaded during the research, and that the hexgrad account has been secured. The researcher did not upload the modified weights he created, but demonstrated they function as intended.

The Hugging Face model page for Kokoro-82M lists 13,092,988 downloads last month and describes the weights as Apache-licensed.

Tampered model weights are part of a wider AI supply chain problem; Microsoft recently warned of post-training supply chain attacks on coding agents. New cases are tracked in our AI Security hub.

Sources

  • One Click From 180 Million Downloads: Taking Over the Hugging Face Account Behind an AI Model (Alon Hertz, Medium)
  • hexgrad/Kokoro-82M · Hugging Face
Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Editor's picks

  1. GAO Report: ChatGPT, Claude, Gemini and Grok Produced Viable Attack Plans After Safeguards Were Bypassed
  2. ResolveAI SRE-1: A 27B Model Trained to Find the Root Cause of Production Incidents
  3. Reson8 Rumble-2-Flash: Open-Weight 0.5B Speech Recognition Model Covers 40+ Languages
  4. Pika Enact Lets Creators Direct AI Video Actors and Cameras in Real Time
  5. Google Stitch Loop: An Always-On Product Agent That Fixes Security, Speed and Accessibility via Pull Requests