Researchers took over the Hugging Face account that publishes Kokoro, the open-weight text-to-speech model with nearly 13 million monthly downloads. Kokoro shows up in voice-agent and on-device speech tooling from Google, NVIDIA, Intel, AMD and AWS. The trick was simple: the original publisher’s domain was listed in a public DONATE.md file, so they registered it and reset the account password. Kokoro hasn’t been updated since April 2025.
Then they showed a quieter attack. They fine-tuned the weights in 87 seconds on a gaming GPU so the model said something different from the input text, with the same code, voice file and config. They tested this only locally and never uploaded the poisoned weights. Because only weights change, there’s nothing for malware scanners to flag, and the researchers say Hugging Face wouldn’t have marked an update like that as suspicious.
My take
Models are dependencies now, and a pinned commit hash is your friend. If your local AI setup pulls “latest” from a publisher account nobody’s watching, you’re trusting whoever controls that email domain next. Until model hubs verify publishers, this is a supply-chain gap. I wrote up a quick how-to on pinning models by revision in this same batch.
God Bless, and Needs Garlic!