InstructLab Vulnerability: Remote Code Execution via Malicious HuggingFace Models

The National Vulnerability Database has identified a critical flaw, CVE-2026-6859, within InstructLab. The linux_train.py script contains a hardcoded trust_remote_code=True setting when loading models from HuggingFace. This configuration allows a remote attacker to execute arbitrary Python code on a user’s system by tricking them into training, downloading, or generating content with a specially crafted, malicious model hosted on the HuggingFace Hub. This vulnerability carries a CVSS score of 8.8 (HIGH) and could lead to a complete system compromise.

This exploit leverages the trust placed in model repositories. By embedding malicious code within a seemingly legitimate model, an attacker can bypass standard security checks. The attack vector is straightforward: convince a user to execute a command that pulls the compromised model, thereby triggering the arbitrary code execution. Defenders must be aware that any system interacting with HuggingFace models via InstructLab is potentially at risk if users are susceptible to social engineering tactics.