skills/nvidia/osmo/osmo-user/Gen Agent Trust Hub

osmo-user

Pass

Audited by Gen Agent Trust Hub on Oct 9, 2026

Risk Level: SAFE
Full Analysis
  • [PROMPT_INJECTION]: The skill architecture uses a 'workflow-expert' subagent to handle complex setup and submission tasks. While a static detector flagged this for potential concealment, the design specifically returns control to the main agent to provide live status updates to the user, thereby increasing transparency rather than hiding actions.
  • [CREDENTIALS_UNSAFE]: The skill includes comprehensive guidelines for handling credentials, including NGC API keys and S3 tokens. It explicitly instructs the agent never to print secret values, never to read token files into the conversation, and to prefer secure environment variables over user prompts.
  • [EXTERNAL_DOWNLOADS]: The skill fetches workflow examples and documentation from the official NVIDIA OSMO repository on GitHub. These resources are used to help users scaffold their own workflows and do not involve execution of untrusted code on the host system.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes workflow logs and events to summarize state and provide troubleshooting advice. This represents a potential indirect prompt injection surface; however, the skill employs subagents and line-count limits to summarize findings, which serves as a mitigation against large-scale injection attempts via log output.
  • [COMMAND_EXECUTION]: The skill's primary purpose is to execute the osmo CLI for workflow orchestration. The instructions strictly limit this to user-level compute operations and explicitly forbid administrative or server-side configuration changes, maintaining a principle of least privilege.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 9, 2026, 05:13 PM