Skip to main content

Kokoro TTS Hugging Face Account Hijacked Through a Lapsed Domain

Security researcher Alon Hertz and his team registered the lapsed hexgrad.com domain, used it to reset the password of the Hugging Face account that publishes Kokoro, an open-weight text-to-speech model, and took control. They fine-tuned the weights in 87 seconds so the model speaks different words than its input, but did not upload them. Hugging Face had not responded at publication.

On this page

How a lapsed domain unlocked the account

Kokoro is an 82-million-parameter open-weight text-to-speech model published on Hugging Face by an account called hexgrad. It supports several languages, ships more than 50 voice presets, and is one of the most widely used local speech models in the open ecosystem. Its developer has not pushed an update since April 2025, yet the model still draws nearly 13 million monthly downloads, according to BleepingComputer.

That silence is what made the takeover possible. The project's public DONATE.md file listed the hexgrad.com domain. Over the weekend before BleepingComputer's 17 November report, a team led by security researcher Alon Hertz registered that lapsed domain, used it to request a password reset on the publisher's Hugging Face account, and took control of it. Hertz says no phishing, zero-day or breach of Hugging Face was involved: one domain registration and one password reset were enough.

87 seconds to change what the model says

The researchers then showed what a malicious "update" could carry. On a gaming GPU, they fine-tuned Kokoro's weights in 87 seconds so that the model speaks different words from the text it is given. Their demonstration fed the line "Here is the cake recipe" to both the original and modified checkpoints through the same, unmodified inference code; the tampered model said omelette instead of cake.

Hertz stresses that nothing else changed: the same input text, voice file, configuration and application code. Because only floating-point weights differ, there is no malicious code, no string and no signature for a malware scanner to catch. File sizes and formats stay the same. The team says it tested the altered model locally and never uploaded it, because that would have put production systems at risk. They repeated the idea on a small Qwen2.5 text model using a hand-built dataset, and report that the modified model answered cake-making questions with an omelette recipe in 80 of 81 cases.

Why the blast radius is so large

Hertz's write-up counts roughly 180 million all-time downloads across Kokoro's distribution channels, about 143.5 million of them from the main Hugging Face repository, and 59 open-source projects that integrate it. He lists Google's AI Edge Gallery, NVIDIA's NeMo voice-agent stack, AMD's GAIA, Intel's OpenVINO and AWS deployment guides among the places it appears. Many of those integrations fetch weights automatically, so any pipeline set to pull the latest revision would have taken a poisoned release without anyone touching its code.

The broader complaint is about platform trust. "Hugging Face does not vet publishers," Hertz writes, arguing that the platform does not check whether an account is still held by its original owner. He reported the issue and received no reply, and Hugging Face had not answered BleepingComputer's request for comment at publication.

What local AI users should take from it

This is the uncomfortable side of running models on your own hardware. Local inference keeps your prompts private, but you still trust whoever published the weights, and an abandoned account can quietly change hands. Three habits blunt the risk. Pin a specific model revision rather than "main" or "latest". Record and verify a SHA-256 hash for every file you download. Prefer mirrors or converted builds you can audit at a fixed commit.

CuriousLM's own read-aloud voice uses Kokoro, so we checked how the app gets it. It downloads an ONNX conversion from the onnx-community organisation, pinned to a fixed commit, and verifies every file against a SHA-256 hash before use. A new upload to the hexgrad account would not reach it. Pinning does not prove a model was clean when it was pinned, though, and the case Hertz makes is that weight-level tampering leaves nothing for a reviewer to read. Ownership checks for dormant publisher accounts are the fix only the platform can ship.

CuriousLM runs supported AI models locally on your device. Try CuriousLM.