Researchers Hijack Hugging Face Account for Kokoro AI Model and Demonstrate Stealthy Weight Poisoning

· Original article ↗

Summary

Researchers say they took over the Kokoro model’s Hugging Face account by registering its listed domain and resetting the password. They demonstrated locally that modified weights could change speech output without altering application code.

Key points

  • Researchers registered the hexgrad[.]com domain listed in Kokoro’s public DONATE.md file and used it to reset the model publisher’s Hugging Face password.
  • They modified Kokoro’s model weights in 87 seconds using a gaming GPU, changing generated audio without modifying the inference application code.
  • The modified model was tested locally; researchers say they did not upload the altered weights to Hugging Face.
  • The researchers warned that routine updates from compromised publisher accounts could distribute poisoned weights to systems that automatically fetch new revisions.
  • The article reports that Kokoro had nearly 13 million monthly downloads and that its developer had not updated the project since April 2025.
  • The researchers reported the issue to Hugging Face, which had not responded at the time of publication.

Article Details

Event Type
Researchers demonstrated an account takeover caused by a lack of ownership verification on the Hugging Face platform.
Impact
Researchers registered the domain listed by Kokoro’s original owner and used it to reset the publisher account password. They modified Kokoro’s model weights in a controlled test, but did not upload the altered weights. The article describes potential supply-chain consequences, including data theft or code execution, but reports no affected production systems.

People

Vendors

Products

Related Articles