Hugging Face Model Hijacked Through Expired Domain
A popular Hugging Face AI model was hijacked via an expired domain, revealing a critical supply-chain vulnerability in open-source AI.
The recent hijack of the popular Kokoro text-to-speech model on Hugging Face is a concrete demonstration of a supply-chain risk we've all been ignoring. The attack, which used an expired domain to take over the publisher account, proves that the trust model for open-source AI is fundamentally fragile. Altering neural weights directly is a far stealthier attack than a malicious pickle file, and most teams are not prepared for it.
How did the hijack work? The attack vector was simple and effective. According to the disclosure, security researchers hijacked a popular AI model by registering an expired domain name associated with the developer's email address. With control of the domain, they could receive the password reset email from Hugging Face, giving them full access to the account. This allowed them to modify the model repository for 'Kokoro', a model with nearly 13 million downloads a month.
What's the real risk here? The researchers didn't just replace a file; they demonstrated the ability to modify the model's neural weights. This is the core of the problem. A compromised model could be altered to produce subtle, malicious outputs. Imagine a text-to-speech model that changes the audio for specific phrases or a code generation model that injects backdoors. Traditional malware scanners looking for executable payloads in `pickle` files would miss this entirely. The attack targets the model's logic, not the surrounding code.
This incident highlights a major gap in Hugging Face's platform security. There is no robust verification of publisher identity or ownership. If a publisher's account is compromised, they can push a malicious update to millions of downstream users without any warning. Trusting a model on the platform is currently an act of faith in the operational security of an often-anonymous developer.