RC RANDOM CHAOS

Kokoro AI Model Hijack Exposes Hugging Face Verification Flaws

· via BleepingComputer

Original source

Researchers hijack Hugging Face account for popular Kokoro AI model

BleepingComputer →

Researchers successfully took control of Hugging Face’s account for Kokoro, a text-to-speech model deployed by Google, NVIDIA, AMD, Intel, and AWS. The attack required minimal sophistication: registering the hexgrad.com domain listed in the model’s public documentation and exploiting Hugging Face’s password reset function. Once inside, they demonstrated a particularly stealthy technique—modifying the model’s neural weights to produce audio that contradicts the input text, while keeping all application code unchanged. Because the attack targets model parameters rather than executable code, traditional malware scanners cannot detect it.

Kokoro remains popular despite being abandoned by its developer since April 2025, accumulating nearly 13 million monthly downloads. The researchers proved they could alter the model’s behavior in 87 seconds using a gaming GPU. If the modified weights were released as a routine update, Hugging Face would not flag anything as suspicious—the publisher would appear unchanged, file formats identical, and no malicious code detectable by scanners.

The underlying problem runs deeper than a single account breach. Hugging Face performs no verification of publisher identities, no ongoing validation of account ownership, and no vetting of uploaded models or datasets. This same unvetted content propagates through mirrors, GitHub repositories, package managers, and deployment templates across the entire open-source AI ecosystem. For systems that automatically fetch the latest model revisions, a compromised account becomes a potent distribution vector for supply-chain attacks. The researchers reported the issue to Hugging Face without receiving a response.

Read the full article

Continue reading at BleepingComputer →

This is an AI-generated summary. Read the original for the full story.