Executive brief
picklescan is a security tool used to scan Python pickle files for malicious code, commonly used in machine learning pipelines and platforms like Hugging Face. A vulnerability allows attackers to bypass the tool's security filters entirely by using a specific Python function to hide dangerous commands. This could allow an attacker to execute arbitrary code on systems that rely on picklescan to verify the safety of uploaded models or data.
Technical details
A universal blocklist bypass exists in picklescan versions prior to 1.0.4 due to the omission of 'pkgutil.resolve_name' from the '_unsafe_globals' list. An attacker can craft a malicious pickle file using chained REDUCE calls: the first call uses 'pkgutil.resolve_name' to resolve a blocked function (like 'os.system') from a string, and the second call executes that resolved function. Because the scanner only inspects global imports and not the arguments or results of REDUCE operations, the dangerous function remains invisible to the scanner. This allows for remote code execution (RCE) on any system using picklescan to validate untrusted pickle data. The issue is addressed in version 1.0.4 by adding pkgutil to the blocklist.
Affected products
- picklescan picklescan < 1.0.4
Timeline
- 2026-03-02: advisory: GitHub Security Advisory published
- 2026-06-17: disclosed: NVD publication date