dgx-station-inference
Pass
Audited by Gen Agent Trust Hub on Oct 3, 2026
Risk Level: SAFE
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill references and fetches content from official NVIDIA (NGC) and vLLM documentation and registries. These are well-known, reputable sources for model containers and deployment guidance.
- [COMMAND_EXECUTION]: The skill uses a local script
dgx-assistto manage the inference lifecycle. It includes explicit 'Hard constraints' that forbid the execution of human-provided setup or serve commands from recipes, mitigating the risk of executing untrusted instructions. - [DATA_EXFILTRATION]: Instructions explicitly state that the agent must pass only credential names, never actual secret values, in transcripts or receipts. Additionally, the workflow defaults to localhost binding and requires explicit confirmation for external network exposure.
- [INDIRECT_PROMPT_INJECTION]: The skill processes external recipes and playbooks. It mitigates injection risks by implementing mandatory preflight checks, dry-run approvals, and restricting the execution of embedded shell text or mutable 'latest' tags from processed data.
Audit Metadata