Security researchers have successfully demonstrated a method to hijack the Hugging Face account associated with Kokoro, a widely used open-weight text-to-speech model. By exploiting a lack of ownership verification on the platform, the researchers gained full control of the account, which is utilized in various AI projects supported by major technology firms including Google, NVIDIA, AMD, Intel, and Amazon Web Services. Kokoro remains a high-profile target due to its popularity, currently seeing nearly 13 million monthly downloads.
The takeover was achieved through a simple domain registration process. The researchers identified that the original developer, Hexgrad, had listed a domain in the project’s documentation that was no longer registered. By purchasing the hexgrad.com domain, the researchers were able to initiate a password reset for the associated Hugging Face account, effectively bypassing the platform's security controls. This process granted them the ability to manage the repository and push updates to the model.
Once in control, the researchers demonstrated a sophisticated supply-chain attack vector by modifying the model’s neural weights. Unlike traditional malware, which relies on malicious code or scripts that can be detected by security scanners, this method alters the output of the model without changing the underlying application code or inference pipeline. By fine-tuning the weights, the researchers were able to force the model to produce audio that deviated from the input text while maintaining the same voice and configuration.
The researchers emphasized that this attack is particularly dangerous because it leaves no digital signature for security software to flag. Because the modification occurs at the weight level, the hijacked model would appear as a legitimate update to any system or developer automatically fetching the latest version from Hugging Face. The team confirmed that they performed this manipulation locally using a standard gaming GPU in under 90 seconds, proving that such an attack is low-cost and highly efficient.
The incident highlights a significant vulnerability in the open-source AI ecosystem regarding the lack of vetting for publishers and models. Hugging Face currently does not verify whether an account remains under the control of its original owner, nor does it perform security audits on the datasets or model weights uploaded by users. This creates a scenario where an attacker can quietly take over a trusted account and distribute poisoned updates that propagate through mirrors, package registries, and deployment guides.
While the researchers successfully demonstrated the ability to hijack the account, they confirmed that they did not upload the poisoned weights to the live Hugging Face platform to avoid disrupting production systems that rely on the model. They noted that their test was intended to serve as a proof-of-concept to draw attention to the systemic risks inherent in how AI models are distributed and updated.
The Kokoro project itself has seen limited activity from its original developer since April 2025, with the associated GitHub repository and PyPI package remaining largely stagnant. Despite this lack of recent maintenance, the model’s integration into various developer tools and AI-agent workflows ensures that it remains a critical component in the supply chain for many organizations. The researchers warned that even without a formal account takeover, the current model of trust allows for the potential of "sleeper" accounts that could push malicious updates at any time.
The researchers reported these findings to Hugging Face but stated they have not received a response. As of now, Hugging Face has not publicly addressed the specific security concerns raised by this incident or indicated if they plan to implement stricter ownership verification or model auditing processes. The vulnerability remains a point of concern for developers who integrate third-party AI models into their infrastructure without independent verification of the source material.





