Researchers took control of the Hugging Face account behind the Kokoro AI model, which records nearly 13 million downloads a month on the platform. In a separate local test, they showed that Kokoro could speak words that differed from the supplied text. The team kept the altered model off the public repository.
Kokoro still draws millions of downloads
Kokoro converts written text into speech in multiple languages and offers more than 50 preset voices. Developer tools from Google, NVIDIA and Intel support the model for uses ranging from voice agents to speech generation on a device. AMD and Amazon Web Services have also published guides showing developers how to use Kokoro in AI projects.
That interest continued long after its developer, Hexgrad, last updated the model in April 2025. Activity on the project’s GitHub repository and its Python package also went quiet around that time. Meanwhile, developers could still download Kokoro from the same publisher account and had little reason to expect that someone else might control its next release.
Local test changes what listeners hear
Security researcher Alon Hertz said his team gained access to the publisher account after the original owner lost control of a domain tied to account recovery. Once inside, the researchers could have published a new revision under Hexgrad’s existing identity. To study what such an update could mean for listeners, they tested a modified copy of the Kokoro AI model locally.
The researchers changed the model’s weights, the numerical values it uses to turn text into speech. They then supplied the same text, selected the same voice and ran the same application code. This time, the voice spoke different words. According to Hertz, a conventional malware scan would have little to flag because the change involved no malicious application code.
An altered release could reach downstream users
For developers who automatically fetch the latest model revision, the risk starts with an otherwise routine update. The download could come from the familiar publisher, arrive through the usual channel and use the expected file format. A listener could then hear something different from what the application displayed on screen.
Hertz also pointed to the way model files circulate after publication. Mirrors, converted copies and package registries can carry a single revision beyond the original repository. He reported the account ownership issue to Hugging Face and said he had received no response by November 17.


0 responses to “Kokoro AI model account takeover exposes update risk”