Malicious AI Models on Hugging Face Backdoor Users’ Machines
At least 100 malicious AI models have been discovered on Hugging Face, with some capable of executing code on victims' machines to create persistent backdoors for attackers.
Published on April 14, 2025
Cybersecurity researchers have confirmed that at least 100 malicious AI machine learning models were found on the Hugging Face platform. Some of these models have the capability to execute code on users' machines, effectively providing attackers with a persistent backdoor.
Recent investigations, including efforts reported on February 10, 2025, highlighted that these malicious models exploited vulnerabilities in serialization techniques such as the Pickle format. On March 5, 2025, additional security measures were bolstered when Hugging Face partnered with JFrog to integrate a more robust scanning solution to reduce false positives and enhance threat detection, further underlining the ongoing battle against cyber threats in AI platforms.