Kokoro AI model: researchers claim Hugging Face account hijack
BleepingComputer reports researchers claim they hijacked the Kokoro Hugging Face account; users should review automatic model updates.
- Published
- Ecosystem
- AI developer toolsOpen source
- Exploitation
- No, the researchers say they never uploaded the modified weights to Hugging Face, and the source reports no malicious use
What happened
BleepingComputer reports that security researcher Alon Hertz and his team claim they gained control of the Hugging Face account for Kokoro, an open-weight text-to-speech model. This is the researchers’ account of events as relayed by BleepingComputer. Neither Hugging Face nor the model’s developer, Hexgrad, has confirmed it in the source.
According to the article, the researchers say they registered the “hexgrad.com” domain listed in Kokoro’s public DONATE.md file and used it to reset the password on the publisher’s account. BleepingComputer says this happened over the past weekend before publication on 2026-11-17.
The researchers then say they ran a controlled test. According to BleepingComputer, they modified the model’s weights in 87 seconds on a gaming GPU so that it produced audio differing from the input text. They say the application code was unchanged. Hertz says the team tested the modified model locally and never uploaded the altered weights to Hugging Face, because that could put production systems at risk.
The researchers also say that, because only the weights change, there is no malicious code or signature for malware scanners to detect. They say tests from another account showed Hugging Face would not flag such an update as suspicious. Hertz is quoted as saying Hugging Face does not vet publishers or verify that an account is still held by its original owner.
BleepingComputer says Kokoro has nearly 13 million monthly downloads on Hugging Face and is supported by developer tools from Google, NVIDIA and Intel. AMD and Amazon Web Services have published guides on using it in AI-agent workflows. The article says Hexgrad has not pushed an update since April 2025, and that the GitHub repository and the PyPI package also went quiet around then.
Not confirmed in the source:
- Any response from Hugging Face. Hertz says he reported the issue and got no reply, and BleepingComputer says it has also contacted Hugging Face without a response so far.
- Whether the account has been returned to Hexgrad.
- Whether anyone other than these researchers has had access to the account.
- Any CVE, fixed version or patch.
Am I affected?
The source does not report any malicious Kokoro release, so there is no confirmed sign that weights you already downloaded are altered. According to the researcher quoted by BleepingComputer, the supply-chain risk is significant for systems that automatically fetch new revisions of a model published through a compromised account. Based on that, the setups most relevant to this report are those that:
- download Kokoro from Hugging Face and automatically fetch new revisions, or
- rely on the publisher account remaining under its original owner’s control.
Hertz also argues that other Hugging Face publishers could pose a similar risk, not only Kokoro. That is his view, not a confirmed finding about any other account.
What to do now
- Identify where you use Kokoro, including the Hugging Face model, the GitHub repository and the PyPI package, and note which revision each deployment loads.
- Check whether any of your systems automatically fetch new revisions of the model. The researcher says the risk for such systems is significant.
- Treat any new Kokoro revision published after the reported takeover with caution until Hexgrad or Hugging Face say more. This is a precaution, not a step taken from the source.
- Watch for a statement from Hugging Face or Hexgrad. BleepingComputer says it has asked Hugging Face for comment.