// Security Advisory Triage

vLLM 0.31.0 patches DoS in multimodal image processors

GitHub advisories — pip · published 2026-10-24 16:00 UTC
ACTION NEEDED PATCH_NOW SECURITY
WHAT CHANGEDCVE-2026-114055 published 23 October 2026. vLLM versions before 0.31.0 allow denial of service through unvalidated mm_processor_kwargs in multimodal image processors for Isaac, InternVL, and H2OVL models. First patched version is 0.31.0.
WHY IT MATTERSAttackers can crash vLLM instances processing multimodal requests with these models by sending malformed processor arguments, disrupting service availability.
WHAT TO DOUpgrade vLLM to 0.31.0 or later if you serve Isaac, InternVL, or H2OVL models with multimodal input. If you cannot upgrade immediately, verify whether you expose these specific models to untrusted input.
Model-reported confidence 85% (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-gx7p-2j49-hfq4
+summary: vLLM : DoS via unvalidated mm_processor_kwargs in multimodal image processors (Isaac, InternVL, H2OVL, etc.)
+severity: high
+published_at: 2026-10-23T20:58:36Z
+cve_id: CVE-2026-114055
+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 0.92, needed 0.5
triage reasoningvLLM is an AI inference framework in dependency trees of services running LLM workloads. DoS via unvalidated input is actionable.
signalsAI provider SDK/framework affected (vLLM) · DoS vulnerability in multimodal processing · High severity rating · First patched version specified (0.31.0) · Known affected models listed (Isaac, InternVL, H2OVL)
suppression—
triage modelanthropic/claude-haiku-4.5 ($0.00000)
writer modelanthropic/claude-sonnet-4.5 ($0.00000)