CVE-2026-4944
Hardcoded Trust Remote Code in vLLM
Publication date: 2026-05-28
Last updated on: 2026-05-28
Assigner: huntr.dev
Description
Description
CVSS Scores
EPSS Scores
| Probability: | |
| Percentile: |
Meta Information
Affected Vendors & Products
| Vendor | Product | Version / Range |
|---|---|---|
| vllm-project | vllm | 0.14.1 |
Helpful Resources
Exploitability
| CWE ID | Description |
|---|---|
| CWE-22 | The product uses external input to construct a pathname that is intended to identify a file or directory that is located underneath a restricted parent directory, but the product does not properly neutralize special elements within the pathname that can cause the pathname to resolve to a location that is outside of the restricted directory. |
Attack-Flow Graph
AI Powered Q&A
Can you explain this vulnerability to me?
The vulnerability exists in vllm-project/vllm version 0.14.1 where the parameter `trust_remote_code=True` is hardcoded in two model implementation files: `vllm/model_executor/models/nemotron_vl.py` and `vllm/model_executor/models/kimi_k25.py`.
This hardcoding bypasses the user's explicit setting of `--trust-remote-code=False`, which normally would prevent execution of remote code.
As a result, it enables remote code execution through malicious HuggingFace model repositories when loading these specific models.
This issue is an incomplete fix for previous vulnerabilities CVE-2025-66448 and CVE-2026-22807, affecting separate code paths in the model implementation files.
Deployments that load NemotronVL or KimiK25 models are particularly impacted by this vulnerability.
How can this vulnerability impact me? :
This vulnerability can lead to remote code execution on systems running vllm-project/vllm version 0.14.1 when loading NemotronVL or KimiK25 models.
An attacker could exploit malicious HuggingFace model repositories to execute arbitrary code remotely, potentially compromising the affected system.
The impact includes full confidentiality, integrity, and availability loss as indicated by the CVSS score (8.8), meaning attackers could steal data, modify or delete information, or disrupt system operations.
What immediate steps should I take to mitigate this vulnerability?
To mitigate this vulnerability, avoid deploying or loading NemotronVL or KimiK25 models in your vllm-project vllm version 0.14.1 environment, as these models contain the hardcoded `trust_remote_code=True` parameter that enables remote code execution.
Additionally, do not rely on the `--trust-remote-code=False` setting alone, since it is bypassed in these model implementation files. Consider updating to a version of vllm that addresses this issue once available.