The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides commands for ingesting external data that could contain instructions. \n
Ingestion points: Commands like sandbox upload (uploads local files), policy set --policy <file> (loads policy YAML), and profile import --url <URL> (fetches remote profiles) ingest external content into the system. \n
Boundary markers: The skill documentation uses standard delimiters like -- to separate CLI arguments from executed commands and --- to separate metadata from policy content. \n
Capability inventory: The tool supports file writing, remote command execution via sandbox exec and sandbox create, and network service exposure. \n
Sanitization: The skill notes that the gateway performs validation on policies and credentials, and it explicitly advises users to review remote profiles before importing them.\n- [EXTERNAL_DOWNLOADS]: The skill references external documentation and supports fetching configuration profiles from URLs. \n
Evidence: The skill links to official documentation on docs.nvidia.com and provides examples of the profile import --url command which fetches YAML or JSON definitions from remote endpoints. \n
Context: References to docs.nvidia.com are for official vendor documentation. The profile import capability is a documented feature of the CLI tool described in the skill.\n- [COMMAND_EXECUTION]: The skill teaches agents how to execute commands within a managed sandbox environment. \n
Evidence: Commands such as sandbox exec, sandbox connect, and sandbox create are used to run processes inside isolated containers. \n
Security measures: The instructions emphasize using -- to prevent argument injection and suggest reading credentials from environment variables to avoid shell history exposure.

openshell-cli