Back
Hugging FaceNovember 16, 20261 sources

Researchers hijack the Hugging Face account behind the popular Kokoro text-to-speech model

AI Analysis

BleepingComputer reports that researchers gained control of the Kokoro text-to-speech model's account on Hugging Face. The weakness was account ownership verification: the Hub did not adequately confirm who controlled the namespace. As a result, an attacker could have published a modified model under a trusted name.

Kokoro matters because of how widely it is used. It is a small, high-quality TTS model embedded in projects associated with Google, NVIDIA, AMD, Intel and AWS. Many pipelines pull models from the Hub by repository name and often by the latest revision. A hijacked account could therefore push malicious weights or loader code downstream with little friction, especially through serialization formats that run code on load.

The incident fits a pattern of model-hub supply-chain risk, from typosquatted repositories to poisoned pickles. It lands as Hugging Face expands its role as central infrastructure: the same week, it integrated OpenRouter directly into the Hub. More traffic flowing through Hub identifiers makes namespace integrity more important.

Practical guidance follows directly. Pin models to commit hashes rather than branches, prefer safetensors, mirror critical models internally, and verify publisher identity. Watch for Hugging Face's response, such as stronger ownership verification, signed releases or provenance attestations, and whether any malicious revision actually reached users.

Sources
AI Briefing
·Vendors·Curated by AI agents · Updated daily · 2026
Built by Koby Almog