Protect.Computer
NEWS

Researchers Hijack Hugging Face Account Behind Kokoro AI Model

· 2 min read · Malicious byte Got hacked
Researchers Hijack Hugging Face Account Behind Kokoro AI Model

Security researcher Alon Hertz and his team took control of the Hugging Face account that publishes Kokoro, a text-to-speech model used in projects from Google, NVIDIA, AMD, Intel and Amazon Web Services, according to BleepingComputer. They did not break a password. BleepingComputer says they registered the “hexgrad.com” domain listed on Kokoro’s public DONATE.md file, then used it to reset the publisher’s login password, over the weekend before the November 17 report. The model’s developer, Hexgrad, has not pushed an update since April 2025, but Kokoro still has nearly 13 million monthly downloads on Hugging Face, BleepingComputer reports.

The team then showed what a hijacker could do. Using a gaming GPU, they changed Kokoro’s model weights in 87 seconds so the model produced audio that differed from the text it was given. Only the weights changed, so there was no malicious code or signature for a malware scanner to find, the researchers say, and tests from another account showed Hugging Face would not flag anything as suspicious. Hertz says he tested the altered model only locally and never uploaded it, because that could put many production systems at risk. He says he reported the problem to Hugging Face and got no response; BleepingComputer also asked the company for comment and had no reply when it published.

The broader point, in Hertz’s words, is that Hugging Face does not verify whether a publisher account is still held by its original owner. Projects that automatically fetch new revisions of a model are the exposed path, BleepingComputer says. Nothing in the reporting says a malicious Kokoro update was ever published.

How to check if you’re affected

Affected products are projects that automatically download the latest version of the hexgrad/Kokoro-82M model, or of any other model whose publisher account may have lapsed. No poisoned version of Kokoro is reported to exist, so there is no bad version number to look for. The steps below are our suggestions; the sources do not give any.

  • Find where your code or deployment pulls Kokoro (or other open-weight models) from, and whether it fetches “latest” on every start or pins a specific revision.
  • If it follows “latest”, consider pinning to a revision you have already tested, and re-test before moving it.
  • Treat model weights like any other third-party dependency: note which publisher accounts you rely on and whether their contact domains still belong to them.

Sources

Related reading