Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-114053 Overview
CVE-2026-114053 is a remote code execution vulnerability in vLLM versions 0.7.3 through 0.31.0. The server forwards the request-supplied mm_processor_kwargs.code_revision parameter to the Transformers AutoProcessor.from_pretrained function. When trust_remote_code is enabled, clients can select which revision of a model's dynamic processor code is imported. An attacker who controls a revision in the served model repository can send a multimodal request to the unauthenticated /tokenize endpoint to execute arbitrary Python code in the vLLM API process. This flaw is tracked as CWE-94: Improper Control of Generation of Code.
Critical Impact
Unauthenticated attackers can execute arbitrary Python code in the vLLM API process when trust_remote_code is enabled, leading to full host compromise.
Affected Products
- vLLM 0.7.3 through versions before 0.31.0
- Deployments with trust_remote_code enabled
- Multimodal model serving configurations exposing /tokenize
Discovery Timeline
- 2026-10-21 - CVE-2026-114053 published to NVD
- 2026-10-22 - Last updated in NVD database
Technical Details for CVE-2026-114053
Vulnerability Analysis
vLLM is a widely deployed inference and serving engine for large language models. The vulnerability resides in how vLLM handles multimodal processor keyword arguments supplied through HTTP requests. The mm_processor_kwargs dictionary is parsed from client input and passed directly to the Hugging Face Transformers AutoProcessor.from_pretrained loader. One of its accepted keys, code_revision, selects which Git revision of a repository's dynamic Python code is imported into the running process.
When the server is launched with trust_remote_code=True, Transformers imports the Python modules referenced by the selected revision without sandboxing. The /tokenize endpoint accepts these parameters without authentication, so any network-reachable client can trigger the import. If an attacker can author or submit a revision in the served model repository, they can supply Python that runs in the vLLM API worker.
Root Cause
The root cause is unvalidated propagation of a client-controlled field into a code-loading API. vLLM does not strip, pin, or allowlist the code_revision value within mm_processor_kwargs before forwarding it to Transformers. Combined with trust_remote_code semantics, this converts a model configuration parameter into an arbitrary code execution primitive.
Attack Vector
Exploitation requires network access to the vLLM HTTP API and a repository revision the attacker controls. The attacker sends a crafted multimodal request to /tokenize containing mm_processor_kwargs with a malicious code_revision. vLLM loads the specified revision's processor module, executing the attacker-supplied Python with the privileges of the vLLM API process.
# Patch reference from vllm/v1/attention/backends/mla/flashinfer_mla_sparse.py
MultipleOf,
)
from vllm.v1.attention.backends.mla.sparse_utils import (
flat_kv_row_view,
triton_convert_req_index_to_global_index,
triton_filter_and_convert_dcp_index,
)
# Source: https://github.com/vllm-project/vllm/commit/570735520942b6428ae98f1eb442c6d88adae55e
Stop Remote Code Execution Before It Runs
Singularity™ Endpoint detects and blocks exploit attempts at machine speed, including zero-days and novel attack techniques.
Stop Exploit AttemptsDetection Methods for CVE-2026-114053
Indicators of Compromise
- HTTP POST requests to /tokenize containing mm_processor_kwargs with a code_revision field originating from untrusted clients.
- Unexpected outbound network connections, shell processes, or file writes spawned from the vLLM API process.
- Hugging Face cache directories showing newly fetched revisions that do not match the pinned production revision.
Detection Strategies
- Inspect reverse proxy and application logs for mm_processor_kwargs.code_revision values submitted at request time.
- Monitor child-process creation under the Python interpreter hosting vLLM for anomalous commands such as sh, curl, or python -c.
- Alert on access to the /tokenize endpoint from source IPs outside the expected inference client set.
Monitoring Recommendations
- Enable verbose request logging on the vLLM OpenAI-compatible server and forward logs to a central analytics platform.
- Track filesystem changes in the Transformers and Hugging Face Hub cache directories used by the service account.
- Baseline normal outbound destinations from inference hosts and alert on deviations consistent with reverse shells or data staging.
How to Mitigate CVE-2026-114053
Immediate Actions Required
- Upgrade vLLM to version 0.31.0 or later, which removes code_revision from the accepted mm_processor_kwargs fields.
- Disable trust_remote_code for any model that does not strictly require dynamic processor code.
- Place the vLLM API behind an authenticating gateway and restrict network exposure of /tokenize to trusted clients.
Patch Information
The vLLM maintainers released a fix referenced in the GitHub Security Advisory GHSA-h3rc-6mm3-gc2m and the upstream commit. Additional analysis is available in the VulnCheck RCE Advisory. Upgrade to vLLM 0.31.0 or later to receive the fix.
Workarounds
- Launch vLLM without --trust-remote-code and only serve models whose code is vendored and reviewed.
- Pin and validate the model revision at startup, and reject any request-supplied code_revision at the proxy layer.
- Run the vLLM process as an unprivileged user inside a container with egress filtering to limit post-exploitation impact.
# Example reverse-proxy filter to strip client-controlled code_revision
# NGINX location block for /tokenize
location /tokenize {
if ($request_body ~* "code_revision") {
return 400;
}
proxy_pass http://vllm_backend;
}
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

