// Security Advisory Triage
vLLM 0.32.0 patches Ghostscript RCE via untrusted EPS inputs
GitHub advisories — pip · published 2026-10-31 05:45 UTC
WHAT CHANGEDGitHub published CVE-2026-115915 on 28 October 2026, rating it high severity. vLLM versions before 0.32.0 allow untrusted EPS inputs to trigger denial of service or remote code execution through Ghostscript. Version 0.32.0 patches the vulnerability.
WHY IT MATTERSAn attacker who can supply EPS files to a vLLM deployment can execute arbitrary code on the inference server. This affects any vLLM instance that processes user-supplied images or documents without input validation.
WHAT TO DOUpgrade vLLM to 0.32.0 or later. If you cannot upgrade immediately, block EPS file uploads and verify that image processing paths reject EPS MIME types.
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-j82p-2c3h +summary: vLLM: Untrusted EPS inputs can trigger denial of service or Ghostscript RCE +severity: high +published_at: 2026-10-28T21:15:29Z +cve_id: CVE-2026-115915 +vulnerabilities[].package.name: vllm +vulnerabilities[].first_patched_version: 0.32.0
HOW IT WAS JUDGED
Triage signals below are model-reported assertions, preserved verbatim; they are not independent arithmetic verification.
| triage verdict | MATERIAL at 0.92, needed 0.5 |
|---|---|
| triage reasoning | vLLM is a core AI inference runtime in many service dependency trees. RCE via untrusted inputs is actionable. |
| signals | AI provider SDK affected (vLLM is a widely-used LLM inference framework) · known exploitation signal: RCE via Ghostscript · high severity rating with concrete attack vector · patched version available (0.32.0) · new advisory entry with CVE assignment |
| suppression | — |
| triage model | anthropic/claude-haiku-4.5 ($0.00000) |
| writer model | anthropic/claude-sonnet-4.5 ($0.00000) |