Trusting Unsloth’s trust setting could mean executing malicious code
Summary
A high-severity vulnerability (CVE-2026-93348) has been disclosed in Unsloth's model loader, allowing arbitrary Python code execution through a maliciously crafted model configuration. Even with `trust_remote_code=False` enabled, attackers can inject malicious Python statements that execute with the permissions of the loading process, potentially exposing sensitive data and credentials. Updates to Unsloth Zoo and Unsloth are available to address the vulnerability.
IFF Assessment
This vulnerability allows for arbitrary code execution, which is a significant threat to defenders as it can lead to compromise of sensitive data and systems.
Severity
Defender Context
This vulnerability highlights the risks associated with processing untrusted model configurations in AI development environments. Defenders should scrutinize AI dependencies, especially in notebooks, training workers, and container images, and ensure that libraries like Unsloth are updated to patched versions to prevent potential code execution and data exfiltration.