Hugging Face says malicious dataset exposed internal credentials

Hugging Face said a malicious dataset uploaded to its platform exploited a security flaw, allowing attackers to run code on company servers and gain wider access to internal systems. The company has revoked and rotated credentials it said were accessed, while TechCrunch reported Monday that Hugging Face was still assessing whether customer or partner data had been taken.

The incident places one of the better-known AI infrastructure platforms in the center of a broader rise in AI-linked cyber activity. Hugging Face hosts datasets and other resources used by developers and researchers, making the integrity of uploaded content and sandboxed execution environments central to its security model.

In a company blog post, Hugging Face said the uploaded dataset took advantage of a vulnerability to execute malicious code on its servers. That access let the attackers increase their privileges and reach a wider set of internal systems, according to the company.

Hugging Face described the campaign as being run by an autonomous agent framework that carried out many thousands of actions across short-lived sandboxes. The company said the operation used self-migrating command-and-control infrastructure staged on public services, and said the activity matched an “agentic attacker” scenario the security industry has anticipated. Hugging Face said the large language model used in the campaign was not known.

Agentic systems can break tasks into steps and execute them with limited human intervention. In a cyberattack, that can mean probing systems, adapting to errors, moving between environments and maintaining access at a speed and scale that differs from more manual intrusion attempts. Hugging Face’s account indicates that the attackers used those capabilities against a platform designed to process untrusted user-submitted material.

The company said it had invalidated and replaced the compromised credentials it identified. It also urged users to rotate their own access tokens and examine their accounts for unusual activity.

TechCrunch reported that Hugging Face disclosed the breach last week and was still determining whether any data belonging to customers or partners had been stolen. The report said attempts to enter corporate networks through stolen credentials or software weaknesses are common, but the Hugging Face incident illustrates the problem facing platforms whose tools and infrastructure can be abused to reach sensitive information.

The breach follows other cyber incidents disclosed this year. PYMNTS reported last week that Fairlife, the dairy company owned by The Coca-Cola Co., had experienced a ransomware event affecting its systems. Coca-Cola said in a news release that Fairlife activated incident response and business continuity protocols after detecting the issue, was assessing the impact with outside advisers and cybersecurity experts, and had notified law enforcement.

Law enforcement has also flagged AI’s role in online crime. The FBI’s Internet Crime Complaint Center said in April that it received 22,364 internet crime complaints last year that referenced AI, with reported losses of $893 million. In its 2025 Internet Crime Report, the FBI said AI-enabled synthetic content is becoming harder to detect and easier to produce, giving criminal actors another tool for fraud against individuals, companies and financial institutions.

A PYMNTS Intelligence report, “Is That Content Generated by AI or Humans? Hard to Tell,” found that AI-generated material can deceive both people and AI systems, increasing pressure on businesses and regulators to address the risk.

This story draws on original reporting from PYMNTS.