serving-llms-vllm
Pass
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONREMOTE_CODE_EXECUTIONPRIVILEGE_ESCALATIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill provides workflows for batch inference and production API deployment which ingest untrusted text data.
- Ingestion points: prompts.txt (SKILL.md) and API requests to the vLLM server (SKILL.md).
- Boundary markers: None present in code examples to delimit external data from instructions.
- Capability inventory: File system writing (results.jsonl), network serving (vLLM server), and local command execution.
- Sanitization: None present; the skill processes and generates text based on external input without filtering.
- [DYNAMIC_EXECUTION]: The documentation includes the --trust-remote-code flag which allows loading and execution of logic from remote repositories at runtime.
- Evidence: vllm serve MODEL --trust-remote-code in SKILL.md and references/troubleshooting.md.
- [REMOTE_CODE_EXECUTION]: The --trust-remote-code flag enables the execution of remote scripts bundled with model files from external hubs.
- Evidence: vllm serve MODEL --trust-remote-code in SKILL.md.
- [PRIVILEGE_ESCALATION]: Instructions suggest modifying system firewall settings using elevated privileges via sudo.
- Evidence: sudo ufw allow 8000 in references/troubleshooting.md.
Audit Metadata