CVE-2026-22807
vLLM affected by RCE via auto_map dynamic module loading during model initialization
CVSS v3
8.8
HIGH
EPSS Score
0.8 %
exploit probability, as of 2026-10-12
CISA KEV
No
known exploited
Exploitation
none
SSVC status
Get an alert for each new CVE published for vllm. Watch vllm in CVE Watch Free account · no credit card
Affected and fixed versions
Packages from OSV.dev and products from the vendors' CSAF advisories.
version not verifiedVersions not verified: read as OSV.dev and the vendors publish them, not checked against each vendor's release list.
Packages
- vllmPyPIfrom 0.10.1 · fixed in 0.14.0
Vendor advisories
vLLM: vLLM: Arbitrary code execution via untrusted model loading
rhaiis/vllm-tpu-rhel9 as a component of Red Hat AI Inference Server · rhelai3/bootc-aws-cuda-rhel9 as a component of Red Hat Enterprise Linux AI (RHEL AI) 3 · rhelai3/bootc-azure-cuda-rhel9 as a component of Red Hat Enterprise Linux AI (RHEL AI) 3
Red Hat Security Advisory: RHOAI 2.25.9 - Red Hat OpenShift AI
Red Hat Security Advisory: Red Hat AI Inference Server 3.3.5 (CUDA)
Red Hat Security Advisory: Red Hat AI Inference Server 3.3.5 (ROCm)
Red Hat Security Advisory: Red Hat AI Inference Server 3.3.5 (Spyre)
Red Hat Security Advisory: RHOAI 2.25.5 - Red Hat OpenShift AI
5 more advisories
Description
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.
Technical details
- CVSS v3 Vector
- CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H
- Published
- 2026-01-21
- Last Modified
- 2026-01-30
Frequently asked questions
What is 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.
Is CVE-2026-22807 actively exploited?
Active exploitation of CVE-2026-22807 has not been confirmed. Its EPSS score was 0.8% on 2026-10-12, the estimated probability of exploitation in the next 30 days.
What is the CVSS score for CVE-2026-22807?
CVE-2026-22807 has a CVSS v3 base score of 8.8 (HIGH severity), with vector string CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H.
Running vllm? Hear about its next vulnerability
CVE Watch matches each new CVE to the vendors and products you list (vendor and product, not the version) and puts it in your alerts with its KEV and EPSS status.
Already have an account? Open CVE Watch
No credit card required · 50 free checks/month · Free API key
Go further
- Following what attackers exploit right now? The CISA KEV additions of the month, with their patch deadlines.
- Running this software? CVE Watch alerts you to new vulnerabilities in your products.
Related vulnerabilities
Same product first, then the same vendor advisory and the same publication week.