// Security Advisory Triage

vLLM RCE via request-controlled mm_processor_kwargs.code_revision

GitHub advisories — pip · published 2026-10-27 18:07 UTC
ACTION NEEDED PATCH_NOW SECURITYAPI
WHAT CHANGEDGitHub published CVE-2026-114053 on 27 October 2026. vLLM versions before 0.31.0 allow remote code execution through request-controlled mm_processor_kwargs.code_revision parameter. Patched in vLLM 0.31.0.
WHY IT MATTERSAn attacker controlling request parameters can execute arbitrary code on the server. This is a direct RCE vector in a widely-deployed LLM serving framework, exploitable through normal API usage without authentication bypass.
WHAT TO DOUpgrade vLLM to 0.31.0 or later. If you cannot upgrade immediately, block or validate the mm_processor_kwargs.code_revision parameter in all incoming requests until patched.
Model-reported confidence 95% (not measured accuracy) Primary source ↗ written by anthropic/claude-sonnet-4.5
Show the reasoning — raw diff, triage verdict, confidence
RAW DIFF
+ghsa_id: GHSA-h3rc-6mm3-gc2m
+summary: vLLM: Request-controlled mm_processor_kwargs.code_revision allows remote code execution
+severity: high
+published_at: 2026-10-27T15:36:17Z
+cve_id: CVE-2026-114053
+vulnerabilities[].package.name: vllm
+vulnerabilities[].first_patched_version: 0.31.0
HOW IT WAS JUDGED
Triage signals below are model-reported assertions, preserved verbatim; they are not independent arithmetic verification.
triage verdictMATERIAL at 0.95, needed 0.5
triage reasoningRCE in vLLM via request-controlled parameter. AI SDK in active use; platform leads run inference services.
signalsknown exploitation indicator: 'remote code execution' in summary · AI provider SDK/framework: vLLM is an LLM serving framework in dependency trees · request-controlled parameter allows RCE: prompt/input handling vulnerability · high severity rating · patched version available: 0.31.0 · new advisory: first publication
suppression—
triage modelanthropic/claude-haiku-4.5 ($0.00000)
writer modelanthropic/claude-sonnet-4.5 ($0.00000)