Researchers Hijack Kokoro AI Model Account on Hugging Face

Researchers took control of the Kokoro text-to-speech model's Hugging Face account and altered its model weights in a controlled test.

Updated on Nov. 17, 2026 in Artificial Intelligence

Researchers Hijack Kokoro AI Model Account on Hugging Face

Live Poll

Do you trust the AI tools you use to be free of tampering?

Researchers hijacked the Hugging Face account for the Kokoro text-to-speech model, which is used in projects from Google, NVIDIA, AMD, Intel, and Amazon Web Services. They modified the model's weights to produce audio that differed from the input text, without changing application code.

Why it matters

Hugging Face does not verify whether an account is still held by its original owner, which could let an attacker use a hijacked account as a distribution channel for supply-chain attacks. Kokoro has nearly 13 million monthly downloads, so a compromised model could reach a large number of downstream users.

Researchers modified Kokoro's model weights in 87 seconds using a gaming GPU, and the altered model produced audio that differed from the input text. Hugging Face has not responded to the report, and it is not clear whether any downstream user has received the modified weights.

The players

Hugging Face

Hugging Face is an AI platform that hosts machine learning models, datasets, and applications for developers and researchers worldwide.

The details

Researchers registered the hexgrad.com domain listed in Kokoro's DONATE.md file, then used it to perform a password reset on the publisher's Hugging Face account. They did not upload the modified weights to Hugging Face, and the platform did not flag a separate test update from another account as suspicious. Hertz reported the issue to Hugging Face and received no response, and BleepingComputer has also not received a comment.

Timeline

  1. - April 2025: Kokoro's developer last pushed an update, and GitHub and PyPI went silent.

  2. - Over the past weekend: Researchers registered hexgrad.com and performed the password reset.

Under the Hood

The incident follows the broader pattern of software supply-chain attacks, in which compromised package repositories become a delivery route for malicious code. Here the weak point is account ownership on an AI model platform, where a model that many projects depend on can be altered without changes to downstream application code.

Developers who rely on Kokoro in their projects should verify the integrity of model files before deployment, since the account controlling updates may not be held by the original owner. Users of applications built on Kokoro may receive audio output that differs from what the input text specifies if a modified model is pulled into a project.

The takeaway

Verifying who controls a model publisher's account matters as much as checking the model file itself. Developers should treat model updates from any source as unverified until the publisher's identity is confirmed through independent channels.

Further reading

For broader context on how AI models are shared and secured, see our Artificial Intelligence coverage.

Live Poll

Do you trust the AI tools you use to be free of tampering?