Researchers take over Hugging Face account behind widely used Kokoro speech model
A malicious update pushed from a hijacked account would look like a routine release, and the platform has not responded to the warning.
Security researchers gained control of the Hugging Face account behind Kokoro, a widely used open-weight text-to-speech model, by registering the domain listed in the developer's public donation file and resetting the password. Kokoro is embedded in tools from Google, NVIDIA, Intel, AMD and AWS, and it still draws close to 13 million monthly downloads despite no update since April 2025. In a controlled test, the team altered the model's weights in 87 seconds on a consumer GPU so it produced audio differing from the input text, without touching application code or triggering any malware signature. They did not publish the altered weights, but they warned that a malicious update from a hijacked or quietly taken-over account would look like a routine release, since Hugging Face does not verify publisher ownership or vet uploaded models. The researcher said he reported the issue to Hugging Face and received no response, and BleepingComputer has also not yet received a reply from the platform.
Why it matters
Kokoro still draws close to 13 million monthly downloads even though its developer has not released an update since April 2025, so many users are running a model whose publisher account has now been shown to be easy to seize. Hugging Face would not flag a modified release from that account as suspicious, which means users who trust the model by name have little built-in protection if weights are changed. The researchers chose not to publish their altered weights, but the episode shows that the platform's current checks do not confirm who is behind a model.
Matters or hype?
Does this story matter, or is it hype? Vote before you see what everyone else said.