serving-llms-vllm
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFEPRIVILEGE_ESCALATIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [PRIVILEGE_ESCALATION]: The troubleshooting documentation includes administrative shell commands for system configuration.
- Evidence: The guide suggests using
sudo ufw allow 8000to modify firewall rules in references/troubleshooting.md. - [REMOTE_CODE_EXECUTION]: Instructions detail the use of the
--trust-remote-codeflag, which allows the execution of arbitrary Python code from remote repositories. - Evidence: Found in SKILL.md and references/troubleshooting.md as a recommended solution for loading custom models from Hugging Face.
- [EXTERNAL_DOWNLOADS]: The skill requires the installation of multiple third-party packages and model weights from external sources.
- Evidence: Includes commands to install
vllm,locust,autoawq,auto-gptq, andflash-attn. - Evidence: References model downloads from Hugging Face, such as
meta-llama/Llama-3-8B-Instruct. - [INDIRECT_PROMPT_INJECTION]: The skill facilitates the processing of large batches of untrusted prompt data.
- Ingestion points: Prompts are ingested from external files like
prompts.txtand via API endpoints. - Boundary markers: Absent; there are no instructions to implement delimiters or filtering for the processed prompts.
- Capability inventory: The skill facilitating file-writing (e.g.,
results.jsonl) and shell execution via the vLLM CLI. - Sanitization: Absent; the provided workflows do not include steps for validating or sanitizing the content of the processed prompts.
Audit Metadata