AI Stack Alert High CVE-2026-114053 For AI Eng, AppSec, SOC / IR, CISO AI-assisted
vLLM <0.31.0: request-controlled mm_processor_kwargs.code_revision enables remote code execution (CVE-2026-114053)
vLLM versions >= 0.7.3 and < 0.31.0 let an untrusted multimodal request set mm_processor_kwargs.code_revision. When the served model uses dynamic auto_map processor code and --trust-remote-code is enabled, this value selects which revision of the processor Python code Transformers imports. The advisory states that importing that module runs arbitrary code in the vLLM API/renderer process. The fix is in vLLM 0.31.0.
Why it matters
Any client that can reach an affected endpoint can run code in the vLLM server process under these conditions. The /tokenize endpoint is documented as unauthenticated even when --api-key is set, which widens exposure unless a proxy protects it.
Do this
- Upgrade vLLM to 0.31.0 or later.
- Until patched, keep --trust-remote-code disabled for models that do not require it, since the default False setting is not affected.
- Place /tokenize and other unauthenticated routes behind an authenticating proxy or gateway.
- Review served-model repositories for community pull-request refs or untrusted revisions, and restrict which revisions the server can fetch.
- Search API and access logs for multimodal requests that include code_revision inside mm_processor_kwargs.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.46%
Signals
- NoCISA KEV
- 8.1CVSS
- 0.46%EPSS
- 37/100Priority
- 85%Confidence
Evidence
“This lets a client replace the operator-reviewed processor module with code from another revision of the same model repository.”
“Importing a Python module is arbitrary code execution in the vLLM API/renderer process.”
“Consequently, `/tokenize` remains unauthenticated even when `--api-key` is configured, unless an external proxy protects it.”
Sources
Affected / related
vLLM, Hugging Face Transformers, mm_processor_kwargs.code_revision request injection
Published 2026-10-27 15:36 UTC · updated 2026-11-08 02:40 UTC · AI-assisted summary, checked against quoted sources · How we score