The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill manages software installation and updates by fetching installers from official vendor domains (ota.nvidia.com and ota-downloads.nvidia.com). This process is secured by strict requirements for explicit user consent before any download begins. The instructions mandate multi-stage verification, including SHA-512 hash checks and Authenticode signature validation (CN=NVIDIA Corporation), to ensure file integrity and authenticity.
[COMMAND_EXECUTION]: Local shell commands are used for legitimate administrative tasks such as version checking, architecture detection, and launching the application tray process. The skill provides a pre-defined PowerShell launcher and explicitly instructs the agent to avoid self-elevation, silent installation flags, or bypassing Windows User Account Control (UAC) prompts, ensuring user oversight of all privileged operations.
[DATA_EXFILTRATION]: There is no evidence of data exfiltration or unauthorized network activity. The skill communication is strictly confined to the local loopback interface (127.0.0.1) for the MCP gateway. Instructions specifically forbid the agent from dumping environment variables, inspecting credential stores, or accessing sensitive files outside of documented application configuration paths.
[INDIRECT_PROMPT_INJECTION]: The skill exposes a surface for processing local media files (MP4, MP3, etc.) to apply AI audio and video effects. However, the processing is performed by a dedicated external binary (NVIDIA Broadcast) as a signal processing task. The skill does not contain patterns that would allow data content to be interpreted as instructions, and it requires the agent to confirm status changes with the user.

nvidia-broadcast