Insufficient verification mechanisms on Hugging Face enabled researchers to seize control of Kokoro, an open-weight text-to-speech model deployed across infrastructure from Google, NVIDIA, AMD, Intel, and Amazon Web Services. Such an account compromise could serve as a vector for supply-chain attacks, potentially enabling data exfiltration or arbitrary code execution on dependent systems.
The research team demonstrated a particularly concerning variant: they altered the model's neural weights to generate audio output that diverged from the input text, leaving the application code untouched. Because only the weights were modified, traditional malware detection tools would find nothing suspicious—no malicious code, no altered signatures.

We changed what this model says by fine-tuning its neural weights. Same input text. Same voice file. Same configuration. Same inference code — not a single line touched in the application that runs it.
Alon Hertz, security analyst
Kokoro converts text into speech across multiple languages and offers more than 50 voice presets. Though its creator, Hexgrad, has not released updates since April 2025, the model remains widely adopted, accumulating nearly 13 million monthly downloads on Hugging Face. Both the GitHub repository and the PyPI package index ceased receiving updates around the same period.
Integration libraries from Google, NVIDIA, and Intel support Kokoro for voice agent applications and on-device speech synthesis. AMD and Amazon Web Services have each published documentation showing how to incorporate the model into AI-agent workflows.
How the takeover unfolded
Hertz's team gained access to the Kokoro account through a straightforward method: they registered the domain name appearing in the project's public DONATE.md file and leveraged it to reset the account password. The domain in question was hexgrad.com, listed by the original publisher.
In a controlled demonstration over the weekend, the researchers modified Kokoro's weights in 87 seconds using a consumer-grade gaming GPU, producing poisoned model output. They tested the altered version locally through the standard inference pipeline but deliberately refrained from uploading the modified weights to Hugging Face, citing concern for production deployments relying on the model.
Testing from a separate account confirmed that Hugging Face would not flag the modified release as suspicious. The compromised update would retain all characteristics of a legitimate release: the publisher identity would appear unchanged, the delivery mechanism would function normally, file formats and architecture would match expectations, file sizes would align with prior releases, and no malware signatures would be present.
A broader trust problem
Hertz emphasizes that the supply-chain exposure created by compromised publisher accounts distributing altered weights through routine updates to systems with automatic fetching of new versions represents a substantial risk. Kokoro served merely as a prominent case study to illustrate the vulnerability.
The researcher notes that even accounts not actively compromised pose a threat. An attacker could maintain a seemingly legitimate presence for an extended period before injecting malicious updates at a strategic moment.
Hugging Face does not vet publishers. It doesn't verify whether an account is still held by its original, legitimate owner — or by whoever quietly took it over last week. It doesn't vet the models or datasets those accounts upload.
Alon Hertz
Hertz further noted that unvetted weights and datasets propagate through mirrors, converted versions, GitHub releases, package registries, deployment guides, and direct downloads across the broader open-model ecosystem.
Hertz reported the vulnerability to Hugging Face seeking expedited remediation but received no response. Kernelwire contacted Hugging Face for comment but has not yet received a reply.