CVE-2026-27893 vllm-project CVE debrief

Who should care

Users of vLLM, particularly those who use the engine for large language models, should be aware of this vulnerability and take steps to mitigate it. This includes updating to version 0.18.0 or later and ensuring that remote code execution is not enabled. Additionally, users should be cautious when loading sub-components from untrusted sources.

Technical summary

The vulnerability is caused by two model implementation files that hardcode `trust_remote_code=True` when loading sub-components, bypassing the user's explicit `--trust-remote-code=False` security opt-out. This allows for remote code execution via malicious model repositories. The issue was patched in version 0.18.0, which removes the hardcoded `trust_remote_code=True` setting. Users can mitigate this vulnerability by updating to version 0.18.0 or later and ensuring that remote code execution is not enabled.

Defensive priority

High priority should be given to updating to version 0.18.0 or later. Additionally, users should review their current configuration and ensure that remote code execution is not enabled.

Recommended defensive actions

• Update to version 0.18.0 or later
• Review current configuration and ensure remote code execution is not enabled
• Be cautious when loading sub-components from untrusted sources
• Monitor for suspicious activity
• Consider implementing additional security controls

Evidence notes

The vulnerability was introduced in version 0.10.1 and patched in version 0.18.0. The CVSS score is 8.8, indicating high severity. The vulnerability allows for remote code execution via malicious model repositories.

Official resources