ProjectDiscovery discloses AI lab and model supply chain breach via employee credentials

Security firm ProjectDiscovery, by scanning public Git history and leaked databases, found live credentials for employees at companies including NVIDIA, Nutanix, and AllenAI, some with write access to Hugging Face model registries and Weights & Biases; Nutanix's token had remained in public history for nearly two years. The research also found roughly 122,000 HuggingFace account passwords leaked, plus thousands of account credentials on platforms such as Kaggle and Ollama, which could be used to push backdoored models or tamper with model weights under an organization's name. The article notes the current ecosystem lacks weight baseline comparison and registry locking mechanisms, and recommends rotating keys, minimizing permissions, enforcing MFA, and reviewing upload history.