The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from the local NVIDIA App environment, such as application names, driver release notes, and game status, which could theoretically contain malicious instructions if the local environment is compromised.
Ingestion points: Tools such as nvapp_client_get_applications, nvapp_overlay_get_status, nvapp_overlay_get_current_game_info, and nvapp_client_get_driver_status fetch data from the local MCP server into the agent's context.
Boundary markers: While explicit character-based delimiters are not used, the instructions enforce strict logical boundaries by mandating the agent "Report only returned fields" and "Interpret the result literally."
Capability inventory: The skill possesses capabilities to launch local applications (nvapp_client_launch_application), modify game optimization settings, and control gameplay recording features.
Sanitization: The skill mitigates risks by prohibiting the agent from inferring missing state, inventing arguments, or attempting to bypass platform-level connection logic.
[SAFE]: The skill implements proactive credential safety measures. It explicitly instructs the agent to never accept, read, or expose authentication tokens and mandates the use of placeholders like <redacted-token> when displaying protocol details to the user.
[SAFE]: The skill enforces a principle of least privilege by defining a restricted workflow. It specifically prohibits the agent from using tools or arguments outside of a documented subset, such as blocking the use of "Desktop Capture" and limiting filter configurations to "rtx_dvc."
[SAFE]: The instructions regarding the concealment of tokens and limiting error reporting to platform-provided messages (Hermes) are evaluated as benign safety guidelines intended to prevent credential leakage and reduce hallucinations rather than attempts to hide malicious activity from the user.

nvidia-app