cbcvebase.
CVE-2026-22807
published 2026-01-21

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…

PriorityP265critical9.8CVSS 3.1
AVNACLPRNUINSUCHIHAH
EPSS
0.74%
52.3th percentile
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.

Affected

15 ranges
VendorProductVersion rangeFixed in
rhaiivllm-gaudi-rhel9
rhaiivllm-neuron-rhel9
rhaiivllm-tpu-rhel9
rhaiisvllm-cpu-rhel9
rhaiisvllm-cuda-rhel9
rhaiisvllm-neuron-rhel9
rhaiisvllm-rocm-rhel9
rhaiisvllm-spyre-rhel9
rhaiisvllm-tpu-rhel9
rhelai3bootc-gaudi-rhel9
rhoaiodh-vllm-gaudi-rhel9
vllm-projectvllm
vllmvllm>= 0.10.1 < 0.14.00.14.0
vllmvllm>= 0.10.1 < 0.14.00.14.0
vllmvllm>= 0.10.1 < 0.18.00.18.0

Detection & IOCsextracted from sources · hover to see the quote

pathvllm/model_executor/models/nemotron_vl.py
pathvllm/model_executor/models/kimi_k25.py
  • Flag vLLM server startup events where `auto_map` dynamic modules are loaded from a model repo/path without `trust_remote_code` being explicitly enabled — this indicates the ungated code execution path described in CVE-2026-22807.
  • Alert on vLLM processes loading NemotronVL or KimiK25 models, as these model implementation files hardcode `trust_remote_code=True`, bypassing user-configured `--trust-remote-code=False`.
  • Detect arbitrary code execution occurring before any API request handling during vLLM model load phase — exploitation does not require API access and triggers at server startup.
  • Monitor for vLLM instances running versions 0.10.1 through 0.13.x loading models from remote Hugging Face repositories or attacker-influenced local directory paths.
  • ·The `--trust-remote-code=False` flag is ineffective for NemotronVL and KimiK25 models in vLLM 0.14.1, as `trust_remote_code=True` is hardcoded in their model implementation files, making user-configured security settings non-operative for these model types.
  • ·CVE-2026-22807 is fixed in vLLM 0.14.0, but the related incomplete fix (CVE-2026-4944) affects vLLM 0.14.1 via separate hardcoded code paths — upgrading to 0.14.0 alone does not fully remediate the attack surface.
  • ·vLLM's `auto_map` module loading during model resolution is the ungated code execution trigger — any deployment where the model repo/path can be influenced by an external party is at risk regardless of API access controls.

CVSS provenance

nvdv3.19.8CRITICALCVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H
ghsa8.8HIGH
osv8.8HIGH
vendor_redhat8.8HIGH
Stop checking back — get the weekly exploitation signal.

Every Monday: what got weaponized or added to CISA KEV in the last seven days — each CVE cross-linked to its PoC, Nuclei template, and detection rule. Free, one email a week, unsubscribe in one click.