Researchers hijack Hugging Face account for popular Kokoro AI model
Researchers exploited a lack of ownership verification on Hugging Face to hijack the Kokoro AI model account, raising concerns over supply-chain attacks.

A group of researchers took control of the Hugging Face account for the Kokoro AI model, a widely used text-to-speech tool, due to inadequate ownership verification on the platform. This incident highlights potential vulnerabilities in supply-chain security for AI models relied upon by major companies.
The researchers demonstrated a significant security flaw on the Hugging Face platform by hijacking the Kokoro account. Kokoro is a popular text-to-speech model, utilized by organizations such as Google, NVIDIA, AMD, Intel, and Amazon Web Services. By taking control of the account, the researchers showed that an attacker could use it as a distribution channel for supply-chain attacks, which could lead to data theft or code execution.
In their demonstration, the researchers modified the model's weights, allowing it to produce altered audio outputs without any changes to the application code. This means that the model would still function normally in terms of its interface and delivery, making it difficult for traditional malware detection tools to identify any malicious behavior. The researchers noted that only the weights of the model were changed, meaning there was no malicious code that could be flagged by security scanners.
- 1Register DomainDomain used by original owner was registered.
- 2Password ResetPerformed password reset on publisher's account.
- 3Modify Model WeightsChanged model weights in just 87 seconds.
- 4Test Modified ModelFunctionality verified locally without uploading.
- 5Potential Supply-Chain AttackAltered weights could lead to data theft.
Kokoro operates as an open-weight model that converts text into speech, supporting multiple languages and featuring more than 50 voice presets. It has become increasingly popular, with nearly 13 million monthly downloads on Hugging Face. Despite its popularity, the developer, Hexgrad, has not released any updates since April 2025, which raises questions about the model's ongoing support and security.
The researchers were able to take control of the Kokoro account by registering the domain listed in the model's public DONATE.md file and performing a password reset on the publisher’s account. This method allowed them to gain access quickly. They modified Kokoro's model weights in just 87 seconds using a gaming GPU, demonstrating how easily an attacker could alter the model's output.
The researchers conducted tests locally to verify the modified model's functionality but did not upload the altered weights to Hugging Face. They indicated that doing so could jeopardize many systems that depend on the model, as an update would appear legitimate and not raise any flags within the Hugging Face platform.
Despite the hijacking, the publisher's account remained unchanged, which included the file delivery route, format, and architecture. The file sizes were similar, and no malware was present in the release. This situation underscores a significant supply-chain risk, as compromised accounts could distribute altered weights through routine updates, potentially impacting numerous production systems.
The researchers noted that even legitimate accounts pose risks, as they can curate a trustworthy image before pushing a malicious update later. They pointed out that Hugging Face does not vet publishers, nor does it verify whether an account is still held by its original owner or has been taken over. The lack of verification allows unvetted models and datasets to circulate widely across the open-model ecosystem.
The researchers reported their findings to Hugging Face, seeking prompt action on the issue, but they did not receive a response. This lack of engagement raises concerns about the platform's responsiveness to security vulnerabilities.
What it meansResearchers exploited a security flaw in Hugging Face, allowing them to hijack an AI model account and alter its outputs without detection.
Ask your IT team this weekWhat verification processes do we have for third-party models we use?How do we monitor updates to AI models in our systems?
DecisionThe IT team reviews third-party model access and security protocols, and security checks that this is done.