The vLLM inference engine contains a critical security flaw that allows unauthenticated attackers to achieve remote code execution by bypassing user-defined security configurations.

This vulnerability is a protection mechanism failure where internal model implementation files hardcode the trust_remote_code=True parameter. This flaw allows an unauthenticated attacker to execute arbitrary code on the host system by loading a malicious model, effectively overriding the user-specified --trust-remote-code=False security opt-out.

The ability for an attacker to achieve remote code execution poses a severe threat to infrastructure integrity, as it grants full control over the server environment. Given the CVSS score of 8.8, this vulnerability presents a high risk of total system compromise, potential data exfiltration, and unauthorized access to proprietary machine learning models or sensitive environment variables.

Immediate Action: Update the vLLM package to version 0.18.0 or later to ensure the security opt-out mechanism is correctly enforced.

Proactive Monitoring: Monitor server logs for unexpected outbound network connections or processes spawned by the vLLM service that deviate from normal inference workloads.

Compensating Controls: Restrict access to the vLLM API to trusted internal networks and implement strict egress filtering to prevent the inference engine from reaching untrusted model repositories.

Public Exploit Available: No (Exploit_available: false).

Organizations using vLLM for model serving must prioritize upgrading to version 0.18.0 immediately. Because this flaw bypasses explicit security controls, it undermines the trust model of the entire deployment, making the application of the vendor-provided patch the only effective way to restore system security.