Cybersecurity

Vulnerability in Hugging Face Allows AI Model Hijacking and Output Poisoning

Security researchers have uncovered a vulnerability in the Hugging Face platform that allows hijacking of model developers' accounts by taking over their domains, opening the door to supply chain attacks by modifying model weights without detection.

2026-11-17
2 min read
14 views
certi.news Editorial Team
Vulnerability in Hugging Face Allows AI Model Hijacking and Output Poisoning

A team of security researchers has successfully demonstrated a severe security vulnerability in the Hugging Face platform, which allows attackers to take complete control of artificial intelligence model developer accounts. This vulnerability was illustrated by hijacking the account of the Kokoro model, a popular Text-to-Speech model that enjoys wide popularity and is used in major technical projects by companies such as Google, NVIDIA, AMD, Intel, and Amazon Web Services.

How did the hijacking take place?

The researchers, led by Alon Hertz, exploited the lack of strict account ownership verification mechanisms on the platform. The team registered the internet domain that was listed in the donation file of the Kokoro project, and then used this domain to reset the developer's account password on Hugging Face, granting them full access to the account.

What does this mean for users?

The researchers explained that the real danger lies not only in account theft, but in the ability to execute supply-chain attacks. Instead of including traditional malware that can be detected by antivirus software, the researchers modified the "weights" of the neural model itself. This modification causes the model to produce audio outputs different from the input text, while the execution code and model infrastructure remain completely unchanged, making the attack invisible to traditional scanning systems.

Why does this news matter?

This breach highlights a broader trust issue in the open-source model ecosystem. The Hugging Face platform currently does not verify the identity of model publishers or whether an account remains under the control of its original owner. Since many production systems rely on automatic model updates, any malicious update pushed through a compromised account will reach thousands of systems relying on this model directly, without raising any suspicions from current security tools.

The researchers emphasize that the Kokoro model was merely a prominent example, and that any account on the platform could be vulnerable to the same scenario, where attackers can wait for a long time before pushing a malicious model update, making the detection of these attacks extremely difficult in real working environments.

News source
BleepingComputer
Open original source ↗
c
Author

certi.news Editorial Team

In the same category

You may also like

View all news