serving-llms-vllm
Fail
Audited by Gen Agent Trust Hub on Feb 16, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [REMOTE_CODE_EXECUTION] (HIGH): The skill documentation (SKILL.md and troubleshooting.md) explicitly instructs users to use the --trust-remote-code flag when serving models. This enables the execution of arbitrary Python code embedded in model repositories from Hugging Face, creating a direct path for Remote Code Execution (RCE) if a malicious model is loaded.
- [COMMAND_EXECUTION] (MEDIUM): Instructions recommend binding the vLLM server to 0.0.0.0 in both Docker and standalone deployment scenarios. Since the examples use a placeholder api_key of 'EMPTY', this configuration exposes the model's inference API to the network without any security controls, potentially allowing unauthorized access to compute resources.
- [EXTERNAL_DOWNLOADS] (LOW): The skill requires multiple third-party Python packages (vllm, locust, autoawq, auto-gptq, flash-attn) to be installed from PyPI. While these are common libraries, they introduce a supply-chain dependency risk as the skill does not specify pinned versions or integrity checks.
Recommendations
- AI detected serious security threats
Audit Metadata