NewApp Monitoring is liveSign up free
AgentbeamBeta
AI agent vulnerability · vLLM

CVE-2026-22807: vLLM affected by RCE via auto_map dynamic module loading during model initialization

highCVSS 3.1 8.8GHSA-2pc9-4j83-qjmr

vLLM 0.10.1 up to but not including 0.14.0 runs Python code from a model repository's `auto_map` entries at server startup, even when `trust_remote_code` is false. An attacker who controls the model repo or local model path gets code execution on the vLLM host. Upgrade to vLLM 0.14.0.

Published Jan 21, 2026 · Updated Jul 21, 2026 · Analysis Nov 1, 2026

What an attacker can do

During model resolution, vllm/model_executor/models/registry.py loops over the auto_map entries in the model config and calls try_get_class_from_dynamic_module. That helper in vllm/transformers_utils/dynamic_module.py hands off to Transformers' get_class_from_dynamic_module, which imports and executes the referenced module. The code never checked trust_remote_code, the flag meant to stop this.

A malicious model repo can ship a Python file that auto_map points to. It runs as the vLLM process when the model loads, before any request is handled, so the attacker needs no API access. The CVSS vector requires user interaction: someone must point vLLM at the attacker's model, such as a Hugging Face repo name from a tutorial, a shared directory or a mirrored model. Code runs with the server's rights, which on GPU hosts often means access to model weights, API keys and cloud credentials.

Who is affected

The vllm pip package from 0.10.1 up to 0.14.0 (exclusive). The advisory says the problem exists wherever the registry resolves auto_map without checking trust_remote_code. Check the installed version:

pip show vllm

Also review which model names or paths your deployments load, including those set in orchestration configs and launch scripts.

vllm

Ecosystem
PyPI
Affected
>= 0.10.1, < 0.14.0
Fixed in
0.14.0
Check
pip show vllm

How to fix it

Upgrade to vLLM 0.14.0:

pip install --upgrade "vllm>=0.14.0"

The fix (pull request 32194) adds a trust_remote_code argument to try_get_class_from_dynamic_module and handles it, so remote module code is not loaded unless you trust it. Red Hat has published errata for its builds. The advisory gives no workaround; until you upgrade, load only models from sources you control and pin them to a reviewed revision.

How to tell if you were hit

Review the model repos and local directories your servers loaded. Look for auto_map in config.json and for the Python files it references, then read those files for network calls, subprocess use or file writes. On hosts that already loaded an unreviewed model, check for unexpected child processes, outbound connections made at startup and new files written by the vLLM user. Cached copies of remote modules in the Hugging Face modules cache directory show which code was fetched.

Original advisory

From the GitHub Advisory Database (CC BY 4.0).

Summary

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.

---

Impact

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.

---

Affected Versions

All versions where vllm/model_executor/models/registry.py resolves auto_map entries with try_get_class_from_dynamic_module without checking `trust_remote_code` (at least current main).

---

Details

During model resolution, vLLM unconditionally iterates auto_map entries from the model config and calls try_get_class_from_dynamic_module, which delegates to Transformers’ get_class_from_dynamic_module and executes the module code.

This occurs even when trust_remote_code is false, allowing a malicious model repo to embed code in a referenced module and have it executed during initialization.

Relevant code

  • vllm/model_executor/models/registry.py:856 — auto_map resolution
  • vllm/transformers_utils/dynamic_module.py:13 — delegates to get_class_from_dynamic_module, which executes code

---

Fixes

  • https://github.com/vllm-project/vllm/pull/32194

Credits

Reported by bugbunny.ai

Credited to zaddy6, arthurgervais, DarkLight1337, russellb.

Questions

Does setting trust_remote_code=false protect me on affected versions?

No. That is the flaw: vLLM resolved auto_map modules and executed them regardless of that flag.

Does the attacker need to send requests to my vLLM server?

No. The code runs during model initialization, before request handling. The attacker only needs the server to load a model repo or path they control.

Which version fixes CVE-2026-22807?

vLLM 0.14.0. The affected range is 0.10.1 up to but not including 0.14.0.

References

How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology

More vLLM vulnerabilities

All vLLM entries