News Reported – not yet confirmed Security news
Kokoro model account takeover on Hugging Face: a lapsed domain was all it took, BleepingComputer reports
A research team led by Alon Hertz registered the domain listed by the author of the Kokoro-82M text-to-speech model and used it to reset the password on the Hugging Face account behind it, according to BleepingComputer. The model page shows nearly 13 million downloads a month. The researchers did not upload any modified weights, the report says.
Key points
- The researchers registered hexgrad.com, the domain listed in the model's DONATE.md file, and used it to reset the author's account password, BleepingComputer says.
- They showed the weights could be altered in 87 seconds on a gaming GPU so the model speaks different words from the text it is given, with no change to application code.
- Hugging Face has not responded to the researchers' report or to the outlet's request for comment, according to the article.
We are reporting this according to BleepingComputer. We could not open the researchers' own write-up on Medium, and Hugging Face has not commented yet, so this story is marked as reported.
What is known
- What: Kokoro-82M is an open-weight text-to-speech model. BleepingComputer says Google, NVIDIA, AMD, Intel and AWS have used it in projects. The model page lists more than 12.9 million downloads in the past month.
- How: the author's account, hexgrad, has not shipped an update since April 2025, the report says. The researchers registered hexgrad.com from the DONATE.md file and reset the account password through an email address on that domain. In their view, Hugging Face does not check whether an account is still controlled by its original owner.
- Impact: an attacker holding the account could push tampered weights to everyone who downloads the model automatically. Hertz says a test upload of a modified model from a different account was not flagged.
What is not known yet
When exactly the takeover happened, whether the account has been secured again, and whether Hugging Face plans to check for lapsed domains.
What to do now
- Pull models from Hugging Face with a pinned revision (the
revisionparameter set to a commit hash), not from themainbranch. - Record the hashes of model files and compare them on every update, for example with our Hash calculator.
- For your own Hugging Face, PyPI and GitHub accounts, make sure the contact email does not sit on a domain you might let expire.
Sources
Spotted a mistake? Let us know and we will correct it openly.