CVE-2026-47155 describes a vulnerability in vLLM, an inference and serving engine for large language models, where revision pinning controls prior to version 0.22.0 do not fully enforce pinned revisions across all loaded artifacts. Specifically, even when a deployment specifies --revision or --code-revision, dynamic code, GGUF files, image processors, retrieval side weights, or same-repository subfolder weights/configurations may be loaded from unpinned or default revisions. This inconsistency leads to a supply-chain integrity issue, as operators may believe they are serving a reviewed and fixed model revision while nested or sibling artifacts outside that revision influence behavior. The issue is resolved in vLLM version 0.22.0.

The vulnerability allows unintended or unreviewed code and model artifacts to be loaded in deployments that rely on revision pinning, potentially leading to integrity violations of the served model. This can result in unauthorized modification of model behavior (high impact on integrity) while confidentiality and availability impacts are low. The CVSS score of 6.5 reflects a medium severity with network attack vector, high attack complexity, no privileges required, and no user interaction needed.

Upgrade vLLM to version 0.22.0 or later, where the revision pinning controls are properly enforced for all model artifacts. No other official remediation or temporary fix is documented. Patch status is confirmed by the vendor advisory stating the issue is fixed in 0.22.0.