Researchers hijack Hugging Face account of Kokoro TTS model via expired hexgrad.com domain, showing model supply-chain risk
Researchers led by security analyst Alon Hertz took control of the Hugging Face publisher account for Kokoro, a popular open-weight text-to-speech model. They registered the hexgrad.com domain listed in the project's DONATE.md file and used it to reset the account password. Kokoro has nearly 13 million monthly downloads. It is used or documented in tooling from Google, NVIDIA, Intel, AMD and Amazon Web Services. In a controlled local test, the team changed the model weights in 87 seconds on a gaming GPU, so the audio no longer matched the input text. The application code was not changed. The modified weights were never uploaded to Hugging Face. Hertz reported the issue to Hugging Face and received no reply. BleepingComputer also got no response. This report is based on the article alone: no web research was done, and no vendor advisory, CVE or evidence of exploitation in the wild is cited.
Detalhes técnicos
- Root cause: Hugging Face does not verify that an account is still held by its original owner. The account's recovery email depended on a domain that had lapsed.
- A poisoned weights update would look like a legitimate release. The publisher, delivery route, file format, architecture and file size would all look normal. Malware scanners would find nothing, because no code or signature changes.
- Systems that automatically fetch new revisions, as well as mirrors, converted copies, GitHub releases and package registries, could spread the altered weights.
Remediação
The article gives no vendor fix. Based on the findings, defenders should pin model revisions by commit hash. They should avoid auto-fetching the latest revision and verify the weights' provenance. Where the publisher has been inactive, they should check whether its contact domains have lapsed.
Linha do tempo
-
Researchers hijack Hugging Face account of Kokoro TTS model via expired hexgrad.com domain, showing model supply-chain risk
ler a análise desta fonte ocultar
Researchers led by security analyst Alon Hertz took control of the Hugging Face publisher account for Kokoro, a popular open-weight text-to-speech model. They registered the hexgrad.com domain listed in the project's DONATE.md file and used it to reset the account password. Kokoro has nearly 13 million monthly downloads. It is used or documented in tooling from Google, NVIDIA, Intel, AMD and Amazon Web Services. In a controlled local test, the team changed the model weights in 87 seconds on a gaming GPU, so the audio no longer matched the input text. The application code was not changed. The modified weights were never uploaded to Hugging Face. Hertz reported the issue to Hugging Face and received no reply. BleepingComputer also got no response. This report is based on the article alone: no web research was done, and no vendor advisory, CVE or evidence of exploitation in the wild is cited.
Technical Details
- Root cause: Hugging Face does not verify that an account is still held by its original owner. The account's recovery email depended on a domain that had lapsed.
- A poisoned weights update would look like a legitimate release. The publisher, delivery route, file format, architecture and file size would all look normal. Malware scanners would find nothing, because no code or signature changes.
- Systems that automatically fetch new revisions, as well as mirrors, converted copies, GitHub releases and package registries, could spread the altered weights.
Remediation
The article gives no vendor fix. Based on the findings, defenders should pin model revisions by commit hash. They should avoid auto-fetching the latest revision and verify the weights' provenance. Where the publisher has been inactive, they should check whether its contact domains have lapsed.