Pillar Security disclosed a critical vulnerability in Unsloth’s model loader, enabling attackers to execute arbitrary Python code through a maliciously crafted model configuration.
The vulnerability or weakness
The vulnerability, tracked as CVE-2026-93348 with a CVSS score of 8.6, lies in Unsloth’s model architecture identification process, which constructs import statements based on unvalidated model-type values.
| CVE | CVSS | Severity | In CISA KEV | KEV due date |
|---|---|---|---|---|
| CVE-2026-93348 | 8.6 | HIGH | No | - |
CVE data: NVD (nvd.nist.gov) and the CISA Known Exploited Vulnerabilities catalog.
Earlier attacks and coverage
No earlier related attack in our archive yet.
Business impact
Organizations may face significant risks, including exposure of sensitive data, model artifacts, and cloud credentials, leading to potential financial losses and operational disruptions.
Industries affected
This issue affects a broad range of industries, including tech, finance, healthcare, and any sector relying on AI models and frameworks.
Users affected
AI developers and users are particularly affected, as the vulnerability can be exploited even when the security setting trustremotecode=False is enabled.
