Security researchers have demonstrated a low-effort supply-chain attack on the Hugging Face ecosystem by hijacking the account responsible for Kokoro, a widely adopted open-weight text-to-speech model. By simply registering an expired domain listed in the project's public files, the team reset the publisher's password and took full administrative control, proving that ownership verification on the platform is virtually nonexistent. This incident highlights a dangerous vulnerability for the nearly 13 million monthly downloads the model receives, affecting integrations with major tech players like Google, NVIDIA, Intel, AMD, and Amazon Web Services.

The Attack Vector: Domain Registration

The hijack was executed without sophisticated hacking techniques. Researcher Alon Hertz and his team identified that the original developer, Hexgrad, had not pushed updates since April 2025. They located the domain 'hexgrad.com' in the project's public DONATE.md file, registered it, and triggered a standard password reset. This allowed them to seize the publisher account, effectively impersonating the original creator of a model that supports multiple languages and over 50 voice presets.

Stealthy Weight Manipulation

Rather than injecting malicious code, the researchers demonstrated a more insidious threat: modifying the model's neural weights. In a controlled test, they altered Kokoro's weights in just 87 seconds using a gaming GPU, causing the model to produce audio that mismatched the input text. "We changed what this model says by fine-tuning its neural weights. Same input text. Same voice file. Same configuration. Same inference code โ€” not a single line touched in the application that runs it," said security researcher Alon Hertz. Because the change occurred within the weights themselves, the application code remained untouched, leaving no signature for traditional malware scanners to detect. The modified model would appear identical in file size, format, and architecture to legitimate releases.

Ecosystem Trust Deficit

Hertz noted that Hugging Face does not vet publishers or verify if an account is held by its original owner. This lack of oversight means that even non-hijacked accounts could eventually push malicious updates after building trust. "The same unvetted weights and datasets flow through mirrors and converted versions, GitHub releases, package registries, deployment guides, and direct downloads across the entire open-model ecosystem," added the researcher. Despite reporting the vulnerability, the researchers received no response from Hugging Face, and BleepingComputer also awaits comment.

Key Takeaways

  • Domain expiry is a critical attack surface for open-source AI projects with inactive maintainers.
  • Weight-only modifications bypass code-based security scans, creating stealthy supply-chain risks.
  • Major tech firms rely on unverified third-party models for production workflows without apparent safeguards.
  • Hugging Face currently lacks proactive ownership verification for publisher accounts.

The Bottom Line

The ease with which researchers hijacked the Kokoro account exposes a fundamental fragility in the open-source AI supply chain. Hugging Face's failure to verify publisher ownership or respond to critical security reports suggests that the platform prioritizes accessibility over security, leaving users vulnerable to silent, undetectable model poisoning.