Critical-Zero Day Vulnerabilities in PickleScan Expose PyTorch Users to Software Supply Chain Attacks
JFrog Ltd. (Nasdaq: JFROG), the Liquid Software company and creators of the JFrog Software Supply Chain Platform, today announced the discovery of three zero-day vulnerabilities (each rated CVSS 9.3) in PickleScan, a widely used industry-standard tool for scanning ML models and ensuring they contain no malicious content. The flaws enable attackers to evade PickleScan’s malware detection and potentially execute a large-scale supply chain attack by distributing malicious ML models that conceal undetectable malicious code, putting developers and organisations at risk.
The vulnerabilities come as the Open Worldwide Application Security Project (OWASP) Top 10 report ranks software supply chain attacks as the third most significant security risk of 2025, illustrating how rapidly AI and ML ecosystems are becoming high-value vectors for threat actors.
The potential impact of these vulnerabilities is highlighted by Hugging Face, hosting over 200,000 models, relying on PickleScan to vet these uploads. If attackers weaponise these vulnerabilities, they could upload malicious models that evade detection and spread widely across the AI ecosystem.
How the Attack Works
The vulnerabilities stem from inconsistencies between how PickleScan evaluates model files and how PyTorch actually loads them. JFrog researchers identified three exploitation paths:
- File Extension Bypass (CVE-2025-10155): Attackers can disguise a malicious pickle file by renaming it to a PyTorch-related extension (e.g., from .pkl to .bin). A flaw in PickleScan’s file type detection allows the scanner to miss the malicious content, while PyTorch loads and executes the code by checking the file’s contents, not its extension.
- CRC Bypass in ZIP Archives (CVE-2025-10156): By introducing CRC errors into zipped model files, attackers can force PickleScan to fail scanning entirely. PyTorch, however, often ignores CRC mismatches, allowing the malicious code to be loaded without detection.
- Unsafe Globals Bypass (CVE-2025-10157): PickleScan blacklists dangerous imports like asyncio, but attackers can exploit subclassing methods to downgrade warnings from “Dangerous” to “Suspicious”, resulting in full arbitrary code execution on model load.
“As AI libraries like PyTorch grow more complex by the day, introducing new features, model formats, and execution pathways faster than security scanning tools can adapt, the widening gap between innovation and protection leaves organisations exposed to emerging threats that conventional tools simply weren’t designed to anticipate,” said JFrog Security Researcher, David Cohen. “Closing this gap requires a research-backed security proxy for AI models, continuously informed by experts who think like both attackers and defenders. By actively analysing new models, tracking library updates, and uncovering novel exploitation techniques, this approach delivers adaptive, intelligence-driven protection against the vulnerabilities that matter most.”
Widespread Impact
JFrog’s security research reveals systemic AI security issues:
- Overreliance on a single scanning tool creates a vulnerable chokepoint
- File-handling inconsistencies expose blind spots
- Even 1-2 vulnerabilities in public repositories hosting millions of models significantly increase the risk of large-scale supply chain attacks
Protection Measures
Organisations and developers must immediately adopt a layered defence strategy beyond simple blacklist matching:
- Update PickleScan: Users should update PickleScan to version 0.0.31, where the maintainers have fixed all disclosed issues.
- Move to Safer Formats: Organisations should restrict the use of unsafe ML model types (such as Pickle) and transition to secure formats like Safetensors, which do not permit arbitrary code execution.
- Layered Security: Do not rely only on a single scanner. Implement isolation measures like a Sandbox and use a Secure model proxy (like JFrog Catalog) for additional protection.
- Automated Scanning Protocol: Any failed scan should automatically quarantine or block the model from being used or distributed.
The JFrog Security Research Team is dedicated to improving security across the software supply chain, including AI and ML pipelines. Stay up-to-date with our security research at https://research.jfrog.com/ or report potential vulnerabilities you discover at security@jfrog.com.
View the full technical breakdown and protective recommendations:
https://jfrog.com/blog/unveiling-3-zero-day-vulnerabilities-in-picklescan
About JFrog
JFrog Ltd. (Nasdaq: FROG), is on a mission to create a world of software continuously delivered without friction from GenAI-powered developers to the distributed edge in a single solution. The JFrog Software Supply Chain Platform is a system of record for DevOps, DevSecOps and MLOps that powers organisations to build, manage, and distribute software quickly and securely, ensuring it is always available, traceable, governable and tamper-proof. Fueled by JFrog Artifactory as the single source of truth for digital businesses’ assets, JFrog’s cloud-native and hybrid platform is available across major cloud service providers’ marketplaces. Millions of users and 7K+ customers worldwide – including the majority of the Fortune 100 – depend on JFrog solutions to securely embrace digital transformation at scale in the AI era. Learn more at www.jfrog.com or follow us on X @JFrog.