VAITP Dataset

← Back to the dataset

CVE-2026-114053

Improper validation of mm_processor_kwargs.code_revision enables remote code execution via /tokenize.

  • CVSS 9.2
  • CWE-94
  • Design Defects
  • Remote

vLLM from 0.7.3 before 0.31.0 forwards the request-supplied mm_processor_kwargs.code_revision to Transformers AutoProcessor.from_pretrained, letting clients select which revision of a model's dynamic processor code is imported when trust_remote_code is enabled. An attacker controlling a revision in the served model repository can send a multimodal request to the unauthenticated /tokenize endpoint to execute arbitrary Python in the vLLM API process.

CVSS base score
9.2
Published
2026-10-21
OWASP
A08 Software and Data Integrity Failures
Orthogonal defect classification
Interface
Code defect classification
Incorrect Interface
Category
Design Defects
Subcategory
Poorly Designed Access Controls
Accessibility scope
Remote
Impact
Arbitrary Code Execution
Affected component
vLLM
Fixed by upgrading
Yes

Solution

Upgrade vLLM to version 0.31.0 or later.

Vulnerable code sample

import json
from transformers import AutoProcessor

def handle_tokenize(request_body: str):
    data = json.loads(request_body)
    model_name = data.get("model")
    # VULNERABLE: passes user-supplied code_revision to AutoProcessor
    processor = AutoProcessor.from_pretrained(
        model_name,
        trust_remote_code=True,
        **data.get("mm_processor_kwargs", {})
    )
    return processor.tokenize(data.get("text", ""))

Patched code sample

import json
from transformers import AutoProcessor

def handle_tokenize(request_body: str):
    data = json.loads(request_body)
    model_name = data.get("model")
    # FIX: ignore untrusted code_revision argument
    safe_kwargs = {
        k: v for k, v in data.get("mm_processor_kwargs", {}).items()
        if k != "code_revision"
    }
    processor = AutoProcessor.from_pretrained(
        model_name,
        trust_remote_code=True,
        **safe_kwargs
    )
    return processor.tokenize(data.get("text", ""))

Published fix

Public sources link this CVE to the changes below. The code samples above were generated to illustrate the defect; these links lead to the code the project actually changed.

  • [Security] Gate per-request multimodal processor kwargs (#58830) (vllm-project/vllm)
    vllm/entrypoints/chat_utils.py, vllm/entrypoints/generate/base/protocol.py, vllm/entrypoints/launchers/api_server/app_state.py · init_app_state, vllm/entrypoints/launchers/cli_args.py, vllm/entrypoints/launchers/render/app_state.py · init_render_app_state, vllm/entrypoints/pooling/base/serving.py, vllm/entrypoints/pooling/base/serving.py · PoolingBaseServing._init_ctx, vllm/entrypoints/pooling/base/serving.py · PoolingBaseServing.__init__, vllm/entrypoints/pooling/factories.py · init_pooling_state, vllm/renderers/online_renderer.py, vllm/renderers/online_renderer.py · OnlineRenderer.preprocess_chat, vllm/renderers/online_renderer.py · OnlineRenderer.preprocess_cmpl and 1 more
  • [Bugfix][MLA] Fix BLHNC addressing for FlashInfer sparse MLA (#54465) (vllm-project/vllm)
    vllm/models/deepseek_v32/common/kernels.py · fused_norm_rope, vllm/models/deepseek_v32/common/kernels.py · _fused_norm_rope_kernel, vllm/v1/attention/backends/mla/flashinfer_mla_sparse.py, vllm/v1/attention/backends/mla/flashinfer_mla_sparse.py · FlashInferMLASparseImpl.forward_mqa, vllm/v1/attention/backends/mla/sparse_utils.py · triton_filter_and_convert_dcp_index

Cite this entry

@misc{vaitp:cve2026114053,
  title        = {{Improper validation of mm_processor_kwargs.code_revision enables remote code execution via /tokenize.}},
  author       = {Bogaerts, Fr\'ed\'eric and Ivaki, Naghmeh and Fonseca, Jos\'e},
  year         = {2026},
  note         = {VAITP Python Vulnerability Dataset, entry CVE-2026-114053},
  howpublished = {\url{https://netpack.pt/vaitp/vulnerability/CVE-2026-114053/}}
}

Download this entry as JSON Download the whole dataset

Introducing the "VAITP dataset": a specialized repository of Python vulnerabilities and patches, meticulously compiled for the use of the security research community. As Python's prominence grows, understanding and addressing potential security vulnerabilities become crucial. Crafted by and for the cybersecurity community, this dataset offers a valuable resource for researchers, analysts, and developers to analyze and mitigate the security risks associated with Python. Through the comprehensive exploration of vulnerabilities and corresponding patches, the VAITP dataset fosters a safer and more resilient Python ecosystem, encouraging collaborative advancements in programming security.

The supreme art of war is to subdue the enemy without fighting.

Sun Tzu – “The Art of War”

:: Shaping the future through research and ingenuity ::

Before you use the VAITP AI models

Please read and accept the following disclaimer:

By choosing "I understand and accept", you acknowledge that you understand and agree to these terms.