Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

Is my AI tool safe?Inference and model servingvLLM project

Is vLLM safe?

No tool is safe or unsafe on its own: it depends on what it can reach and how it is set up. The record for vLLM: 0 incidents in this center's library and 9 CVEs in its advisory feed (2 critical, 7 high), the newest published Oct 28, 2026. Below: what went wrong, what was fixed, and how to use it safely.

What to watch for

The standard open inference server, and second on the feed by CVE count: pickle over its internal ZeroMQ and NCCL channels and code loaded from model configs, each giving code execution to whoever can reach the service.

How to use it safely

Put it behind an authenticating gateway, load only safetensors models you have checked, and track its advisories like any other internet-facing service.

What it can reach

  • Often exposed to the internet
  • Loads model files that can execute code

Known CVEs

  • CVE-2026-115915 High vLLM passes request images to Pillow without limiting formats, so an EPS file can reach Ghostscript
    Oct 28, 2026 · fixed in 0.32.0
  • CVE-2026-114055 High vLLM denial of service via unvalidated mm_processor_kwargs in multimodal image processors
    Oct 23, 2026 · fixed in 0.31.0
  • CVE-2026-114053 High vLLM remote code execution via a request-chosen mm_processor_kwargs.code_revision
    Oct 20, 2026 · fixed in 0.31.0
  • CVE-2026-115912 High vLLM shared-memory multimodal cache handle forgery reaches pickle.loads()
    Oct 20, 2026 · fixed in 0.32.0
  • CVE-2025-66448 High vLLM remote code execution via model config auto_map
    Dec 2, 2025 · fixed in 0.11.1
  • CVE-2025-62164 High vLLM prompt embeddings deserialization (DoS, potential RCE)
    Nov 20, 2025 · fixed in 0.11.1
  • CVE-2025-47277 Critical vLLM remote code execution via the PyNcclPipe service
    May 20, 2025 · fixed in 0.8.5
  • CVE-2025-30165 High vLLM multi-node cluster remote code execution via ZeroMQ pickle
    May 6, 2025 · fixed in 0.10.0
  • CVE-2025-32444 Critical vLLM remote code execution via pickle over ZeroMQ
    Apr 29, 2025 · fixed in 0.8.5

How many people use it

  • 3.2M PyPI downloads in a month (to Oct 1, 2026) · PyPI Stats
  • 91k GitHub stars · GitHub

A high CVE count can mean heavy scrutiny as much as weak engineering, so read the incidents and advisories, not only the number. The counts are what this center has recorded, not every flaw ever found; usage figures are the vendors' and registries' own published numbers.

Sources

Compare it with the others on the Technology watch list