PatchSiren

PatchSiren cyber security CVE debrief

CVE-2026-4944 vllm-project CVE debrief

CVE-2026-4944 is a HIGH severity vulnerability (CVSS 8.8) in vllm-project/vllm version 0.14.1, published 2026-05-28. The issue involves hardcoded `trust_remote_code=True` parameters in two specific 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 `--trust-remote-code=False` command-line setting, enabling remote code execution when loading malicious HuggingFace model repositories. The vulnerability represents an incomplete fix for two prior CVEs (CVE-2025-66448 and CVE-2026-22807), affecting separate code paths in model implementation files. Deployments specifically loading NemotronVL or KimiK25 models are particularly at risk. The vulnerability is classified under CWE-22 (Path Traversal) according to the secondary source.

Vendor
vllm-project
Product
vllm-project/vllm
CVSS
HIGH 8.8
CISA KEV
Not listed in stored evidence
Original CVE published
2026-05-28
Original CVE updated
2026-05-29
Advisory published
2026-05-28
Advisory updated
2026-05-29

Who should care

Organizations running vllm inference services, particularly those deploying NemotronVL or KimiK25 models; ML platform operators using vllm for production inference; security teams monitoring ML supply chain risks

Technical summary

The vllm inference engine version 0.14.1 contains hardcoded `trust_remote_code=True` assignments in model-specific implementation files for NemotronVL and KimiK25 architectures. When users explicitly disable remote code trust via `--trust-remote-code=False`, this user preference is overridden by the hardcoded values in these specific model loaders. This allows malicious HuggingFace repositories to execute arbitrary code during model loading. The vulnerability persists despite prior fixes for related CVEs, indicating incomplete remediation across all model implementation code paths.

Defensive priority

HIGH

Recommended defensive actions

  • Upgrade vllm to a version containing the complete fix for this vulnerability
  • Audit deployments for use of NemotronVL or KimiK25 models and assess exposure
  • Review and restrict HuggingFace model repository sources to trusted origins
  • Implement network egress controls to prevent unauthorized outbound connections from vllm inference workloads
  • Monitor for anomalous code execution or model loading behavior in vllm deployments
  • Verify that `--trust-remote-code=False` is explicitly set and effective in all vllm deployments

Evidence notes

Vulnerability confirmed through official CVE record and NVD entry. Source references include Huntr bounty platform. CVSS vector: CVSS:3.0/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H. Affected version specifically identified as 0.14.1.

Official resources

2026-05-28