CVE-2026-6859

OVERVIEW CVE-2026-6859 is a critical arbitrary code execution vulnerability in InstructLab's linux_train.py script that stems from hardcoding "trust_remote_code=True" when loading models from HuggingFace. This configuration allows threat actors to execute malicious Python code by distributing crafted models through the public HuggingFace Hub, affecting users who run ilab train, download, or generate commands. SEVERITY The vulnerability carries a CVSS 3.1 score of 8.8 (HIGH) with a network-based attack vector, low complexity, no required privileges, and user interaction needed. The impact is severe across all three dimensions: attackers can achieve confidentiality breaches, system integrity compromise, and availability disruption. Successful exploitation results in complete system compromise with arbitrary code execution in the user's execution context. EXPLOITATION STATUS While not currently listed on CISA's Known Exploited Vulnerabilities catalog, the vulnerability is flagged on the Active Hot List indicating active security monitoring and community attention. The EPSS score of 0.00147 indicates a relatively low probability of exploitation compared to baseline vulnerabilities, suggesting limited active exploitation in the wild at this time. However, the low barrier to entry for creating malicious models and the reliance on social engineering create a meaningful risk for targeted or opportunistic attacks.