Security News supply-chain open-weights authentication
SEC News High

Researchers took over the Kokoro model's Hugging Face account

According to BleepingComputer, researchers registered a domain listed in Kokoro's public DONATE.md file, reset the publisher's password, and then poisoned the model's weights in a local test.

KT Kusal Tharindu
On this page

What happened

According to BleepingComputer, researchers took control of the Hugging Face account that publishes the Kokoro text-to-speech model by registering the hexgrad.com domain listed in the project’s public DONATE.md file and using it to reset the account’s login password.

What is affected

Kokoro is an open-weight text-to-speech model with more than 50 voice presets, and it records nearly 13 million monthly downloads on Hugging Face. Hexgrad, the model’s developer, has not pushed an update since April 2025. The report says developer tools from Google, NVIDIA and Intel support the model, and that AMD and Amazon Web Services (AWS) have published guides showing its use in AI-agent workflows.

Why it matters in production

In a controlled test, the researchers modified Kokoro’s model weights in 87 seconds on a gaming graphics processing unit (GPU), causing the model to produce poisoned output. Because only the weights change, the report says there is no malicious code or signature for malware scanners to detect. Alon Hertz, whose team ran the test, says Hugging Face does not vet publishers, does not verify whether an account is still held by its original owner, and does not vet the models or datasets those accounts upload.

What to do

According to the report, Hertz reported the issue to Hugging Face and received no response, and the cited sources name no fixed version or mitigation step. The team tested the modified model locally and never uploaded the modified weights to Hugging Face.

Sources

  1. BleepingComputer: Researchers hijack Hugging Face account for popular Kokoro AI model
3 posts

Discussion

Spotted an error? Send corrections to [email protected] .

Search Oxelan Signal

Type to search the archive.

Search by title, tag, summary or body text.