serving-llms-vllm
Fail
Audited by Gen Agent Trust Hub on Feb 17, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATION
Full Analysis
- REMOTE_CODE_EXECUTION (HIGH): The skill repeatedly promotes the use of the "--trust-remote-code" flag in SKILL.md and references/troubleshooting.md as a solution for loading custom models. This flag bypasses security sandboxing in the transformers and vLLM libraries, allowing the execution of arbitrary Python code defined within the remote model's repository.
- COMMAND_EXECUTION (MEDIUM): The documentation provides shell commands that modify host security configurations, such as opening firewall ports ("sudo ufw allow 8000") and granting containers full GPU hardware access without specifying security constraints.
- EXTERNAL_DOWNLOADS (LOW): The skill workflows involve installing numerous Python packages (locust, autoawq, auto-gptq, flash-attn) from public registries and fetching model weights from Hugging Face. While typical for ML operations, these represent external code entry points.
- DATA_EXFILTRATION (LOW): The skill configures the LLM server to listen on all network interfaces ("--host 0.0.0.0") and exposes a metrics endpoint on port 9090. Without the implementation of authentication or network access control lists, this configuration risks unauthorized access to model data and internal system performance metrics.
Recommendations
- AI detected serious security threats
Audit Metadata