Security researchers successfully gained control of the Hugging Face account associated with Kokoro, an open-weight text-to-speech model. The method involved registering the 'hexgrad.com' domain, which was listed in Kokoro's public DONATE.md file and used by the original developer, Hexgrad. After registering the domain, the researchers used it to perform a password reset on the publisher's Hugging Face account.
With control of the account, the researchers demonstrated a supply-chain attack by modifying the Kokoro model's weights. This modification caused the model to produce audio that differed from the input text, even though the underlying application code remained unchanged. This type of attack is difficult to detect with traditional malware scanners because no malicious code or signature is introduced into the application itself, only the model's internal parameters are altered.
Kokoro is a popular text-to-speech model supporting multiple languages and over 50 voice presets, with nearly 13 million monthly downloads on Hugging Face. Despite its developer not pushing updates since April 2025, the model is integrated into developer tools and workflows from major tech companies, including Google, NVIDIA, AMD, Intel, and Amazon Web Services. This widespread use highlights the potential impact of such a supply-chain compromise.
The researchers emphasized the stealthy nature of the attack. As researcher Alon Hertz stated, they changed what the model says by fine-tuning its neural weights, keeping the input text, voice file, configuration, and inference code identical. The team tested the modified model locally but did not upload the poisoned weights to Hugging Face to avoid compromising production systems. They also noted that Hugging Face did not flag such modifications as suspicious in their tests.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.
One email a day. Unsubscribe in one click, any time.
Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.
▶ Play today's briefNew every morning, and the back catalogue is archived by date.
Security researchers took control of the Hugging Face account for Kokoro, a widely used text-to-speech AI model, by re-registering the developer's expired domain and resetting the password. This vulnerability allowed them to demonstrate how an attacker could modify the model's weights to produce altered audio output without changing application code, posing a supply-chain risk for projects using the model.