Remote Code Execution Vulnerability in vLLM by vLLM Project
CVE-2026-114053

9.2CRITICAL

Key Information:

Status
Vendor
CVE Published:
21 October 2026

What is CVE-2026-114053?

A vulnerability in vLLM allows attackers to exploit the request-supplied mm_processor_kwargs.code_revision parameter, which forwards to Transformers AutoProcessor.from_pretrained. By manipulating model revisions in the repository, an unauthorized actor can target the /tokenize endpoint, enabling arbitrary Python code execution within the vLLM API. This flaw highlights significant risks associated with the trust_remote_code feature, affecting the integrity and security of systems utilizing this product.

Affected Version(s)

vllm 0.7.3 < 0.31.0

vllm 0.31.0

References

CVSS V4

Score:
9.2
Severity:
CRITICAL
Confidentiality:
High
Integrity:
High
Availability:
High
Attack Vector:
Network
Attack Complexity:
Low
Attack Required:
Physical
Privileges Required:
Undefined
User Interaction:
None

Timeline

  • Vulnerability published

  • Vulnerability Reserved

Credit

adithyan-ak
jperezdealgaba
DarkLight1337
.