Security researchers took control of the Hugging Face account behind Kokoro, a widely used text-to-speech AI model, by registering an abandoned domain and resetting the publisher’s password. Their investigation showed how a compromised account could distribute altered models through trusted channels without introducing malicious code.
The research, led by Alon Hertz, raises concerns about the security of open-weight AI models used in products and developer tools associated with Google, NVIDIA, AMD, Intel, and Amazon Web Services. The researchers did not release a poisoned model, and Hertz says the account has since been secured.
An Abandoned Domain Leads to a Trusted Account
Kokoro-82M is an 82-million-parameter model that converts text into speech, supports multiple languages, and offers more than 50 voices. Its popularity has continued even though the main Hugging Face model repository has not been updated since April 2025.
Hertz’s team noticed that hexgrad.com, a domain referenced in the project’s public documentation, was available for registration. They acquired it and used control of the domain to reset the password for Hexgrad’s Hugging Face account.
No phishing attack, software vulnerability, or intrusion into Hugging Face’s infrastructure was necessary. The researchers obtained access to the existing publisher identity, complete with the account’s history and the apparent legitimacy of its published models.
In his original research, Hertz estimated Kokoro’s footprint at roughly 180 million downloads. He identified its use or support in Google’s AI Edge Gallery, NVIDIA’s NeMo tooling, AMD’s GAIA, Intel’s OpenVINO ecosystem, and other services.
These connections do not mean those companies were compromised. They demonstrate the potential reach of an upstream model if an attacker can publish an update under a trusted account. Applications configured to fetch the newest revision could ingest modified weights without any change to their own code.
Model Weights Changed, Application Code Untouched
To illustrate that risk, the team modified Kokoro’s neural weights on a gaming GPU in 87 seconds. The altered model was tested locally using the normal inference pipeline.
Given text about a cake recipe, the original model generated the expected speech. The modified version instead spoke about an omelette, despite receiving the same input text, voice file, and configuration.
That difference matters because the researchers changed the model’s behavior without inserting malware or editing the application that ran it. The output manipulation was contained in the numerical parameters that govern the model’s responses.
A routine update containing such weights might retain the familiar publisher, repository, file format, and delivery mechanism. Conventional malware scans would not necessarily detect a behavioral change, while a new file hash would merely confirm that the model had been updated, not whether its outputs remained trustworthy.
Hertz’s team also tested the idea with Qwen2.5-0.5B-Instruct, a language model. After fine-tuning it on a specially constructed dataset, the modified model returned an omelette recipe to cake-making questions in 80 of 81 reported test cases. The unmodified version did not produce that exact response in 30 baseline tests.
The Implications for AI Supply Chains
For CISOs, the finding highlights a gap between checking where an AI artifact comes from and verifying what it actually does. A legitimate-looking publisher account and clean malware scan cannot establish that model weights or training data are safe.
Organizations using third-party models should inventory those dependencies, avoid automatically deploying unreviewed updates, and pin production deployments to approved versions. Behavioral regression tests can help identify unexpected changes before revised models reach sensitive workflows, particularly those involving customer communications or automated decisions.
The researchers never uploaded their modified Kokoro weights to Hugging Face. In his disclosure, Hertz said the team was coordinating with the platform, that no data had been leaked, and that the account was secured.
The demonstration nevertheless illustrates how a relatively simple account recovery weakness could become an AI supply chain issue. The threat is not limited to malicious code hidden in a download: it can also involve a legitimate-looking model that quietly produces the wrong answer.
John Kevin Hao is a news and feature writer covering cybersecurity, technology, and business targeted for professional audiences.

