RCE in Vllm

CVE-2026-22807

vLLM is an inference and serving engine for large language models (LLMs). Starting in version 0.10.1 and prior to version 0.14.0, vLLM loads Hugging Face `auto_map` dynamic modules during model resolution without gating on `trust_remote_code`, allowing attacker-controlled Python code in a model repo/path to execute at server startup. An attacker who can influence the model repo/path (local directory or remote Hugging Face repo) can achieve arbitrary code execution on the vLLM host during model load. This happens before any request handling and does not require API access. Version 0.14.0 fixes the issue.

Vulnerability class: RCE (Remote Code Execution)

Published · last modified .

CVSS v3 metric

CVSS v3 base score 8.8 (High). Vector: CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H.

EPSS exploit prediction

EPSS: 0.008 (56.4th percentile), scored .

Very low probability of exploitation in the next 30 days; routine patching cadence is appropriate. 56th percentile — 56.4% of CVEs in the catalogue have a lower EPSS than this one. How to read EPSS.

EPSS trend (30 days)EPSS over the last 30 days for CVE-2026-22807: held from 0.008 to 0.008.

EPSS over last 30 days · oldest: 0.008 · newest: 0.008 · change: -0.000

Affected products

Weakness classification (CWE)

Public proof-of-concept exploits

References

Frequently asked questions

What is CVE-2026-22807?
CVE-2026-22807 is a high-severity vulnerability in Vllm, classified under Code Injection. CVSS score: 8.8/10. Published 2026-01-21.
How severe is CVE-2026-22807?
High severity. CVSS v3 base score is 8.8 out of 10.
Is CVE-2026-22807 known to be exploited?
1 public proof-of-concept repositories are indexed. Not currently listed in the CISA KEV catalog.