Hugging Face confirms breach affected internal datasets and credentials, urges users to take action

Hugging Face confirmed on Friday that a recent cyberattack compromised its internal datasets and service credentials. The breach occurred last week, according to the company, and it is currently investigating whether any customer or partner data was accessed. In a blog post, Hugging Face explained that a dataset uploaded to its platform exploited a security flaw to execute malicious code on its servers, enabling attackers to elevate their privileges and access broader parts of the company's internal infrastructure. The company stated that it has revoked and rotated the stolen credentials and is urging users to do the same with any keys they have stored on the platform. Users are also advised to review their accounts for any unusual activity. Hugging Face claims it has resolved the specific vulnerability that was exploited during the attack. However, the incident highlights the ongoing challenge faced by organizations like Hugging Face when adversaries exploit platforms and tools to access and potentially steal sensitive information. Hugging Face attributed the breach to an external AI agent, which carried out numerous actions across multiple temporary sandboxes. These actions included self-migrating command-and-control operations hosted on publicly accessible services. Despite being contacted by TechCrunch, the company did not provide immediate evidence supporting this assertion. The breach was detected by Hugging Face's own anomaly detection system, which analyzed server logs using an AI model. Initially, the company attempted to use a frontier AI model from a commercial provider to analyze the logs. However, this effort was hindered by the provider’s restrictions. As a result, Hugging Face opted to use its own locally hosted large language model, which allowed for the analysis without uploading sensitive data to third-party servers. Security experts have raised concerns about certain frontier models, such as Anthropic’s Mythos and Fable, which impose strict limitations that can impede cybersecurity investigations. These models have been the subject of disputes with the Trump administration, particularly regarding their potential use in offensive cyber operations. Anthropic was compelled to remove Fable from public availability following the imposition of export controls by the U.S. government. Hugging Face has informed law enforcement authorities about the incident and enlisted cybersecurity forensic specialists to conduct a thorough investigation and evaluate its security measures. There is uncertainty surrounding whether Hugging Face conducted a prior security audit of its systems before launching its services. A representative from the company did not respond to a request for clarification on Monday. The incident raises important questions about the security practices of platforms hosting AI models and datasets, especially given the increasing sophistication of cyber threats targeting such environments.