Researchers report taking control of the Hugging Face publisher account behind Kokoro, a widely downloaded text-to-speech model. Alon Hertz says his team registered an abandoned domain associated with the publisher and used account recovery to gain access.
The researchers separately altered the model’s weights, the numerical data that determines its behavior. In their local demonstration, unchanged application code read the same text but produced different spoken words. They say they never uploaded the modified weights. This is a reported account takeover and controlled demonstration, not evidence of a poisoned public Kokoro release.
Two different trust failures
Hertz’s full account describes control of the hexgrad account, publisher of Kokoro-82M. Losing the domain tied to an account’s recovery route can transfer control without changing the familiar publisher name.
The weight modification demonstrates a different issue. A standard model file can change behavior even when the program loading it remains unchanged. A malware scan can check for known malicious code without establishing that the model will say the intended words.
Hertz reports changing the spoken output from a cake recipe to an omelette recipe after 87 seconds of fine-tuning on a gaming GPU. That is a researcher-reported example, not a measured success rate across arbitrary phrases, voices or languages.
Evidence status: researcher-reported publisher takeover, with locally tested altered weights. BleepingComputer reports the claim and says Hugging Face had not replied to its request for comment. The primary account now says the account has been secured and that coordination with Hugging Face is underway. That recovery claim comes from Hertz, not a reviewed platform incident report.
Who could receive an unwanted revision
The relevant exposure is a workflow that automatically downloads a new revision from a compromised publisher. Existing copies do not change merely because somebody gains control of the account. A pinned, previously obtained artifact and a fresh download of a moving branch are different trust decisions.
The model’s download totals and its appearances in major companies’ tooling show reach. They do not count compromised machines or establish that those companies would automatically deploy a new checkpoint. The reviewed sources identify no downstream deployment of the researchers’ modified model.
Keep model updates explicit and reproducible
For teams consuming Kokoro or other downloaded models:
- Record the exact model revision in each release. Hugging Face’s download documentation supports selecting a revision by full commit hash. Its default download uses the latest revision on
main. - Pin to a previously reviewed revision and retain its expected file hashes. Comparing against a value saved before an unwanted change can detect replacement. Accepting a new hash from the same compromised publisher cannot establish that the new model is trustworthy.
- Review and test weight changes before promotion. For speech systems, compare the audio actually produced with representative input text, including consequential names, numbers and instructions.
- Keep a rollback copy of approved artifacts. A behavioral test can catch a regression without proving that every possible trigger is absent.
For publishers, maintain recovery domains and protect the associated email accounts with strong authentication. Treat recovery routes as production credentials, even when development has paused.
These controls address future update risk. The sources provide no affected package version, malicious release identifier or basis for declaring all current Kokoro installations compromised.
Caveat: The takeover, recovery and behavioral results rely on the researchers’ account. The reviewed reporting supplies no independent reproduction or completed Hugging Face forensic report.