The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill manages sensitive credentials including NVIDIA_API_KEY and NGC_API_KEY. It includes explicit security warnings across multiple files (SKILL.md, references/setup.md, references/asr.md) advising the user to avoid printing, logging, or asking for these keys in chat. It recommends the use of chmod 600 permissions for environment files and the use of credential managers to prevent unauthorized access to these secrets. All network operations for function discovery target the official NVIDIA domain api.nvcf.nvidia.com.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill references several external dependencies including the nvidia-riva-client Python package and various GitHub repositories such as github.com/nvidia-riva/python-clients, github.com/nvidia-riva/cpp-clients, and github.com/NVIDIA-NeMo/NeMo. These resources originate from the official 'nvidia' vendor accounts or organizations and are recognized as trusted well-known services. It also pulls container images from the official NVIDIA NGC registry (nvcr.io). The skill does not perform any unverifiable remote script execution.
[COMMAND_EXECUTION]: The skill provides numerous shell commands for Docker container management, NVIDIA driver verification, and Riva CLI tools (riva-build, riva-deploy). These commands are standard for the intended purpose of deploying and testing machine learning models. The skill recommends best practices for host directory mounting, such as creating directories with specific ownership (chown 1000:1000) rather than using world-writable permissions.
[PERSISTENCE_MECHANISMS]: The skill provides instructions for persisting environment variables like NGC_API_KEY in shell configuration files (~/.bashrc or ~/.zshrc). This is presented as a user-optional step for convenience and is accompanied by a security note regarding the risks of plaintext storage, suggesting more secure alternatives like dedicated .env files with restricted permissions.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection as it processes external audio data for transcription. However, it follows best practices by instructing the agent to use specific gRPC and HTTP client protocols with defined parameters, minimizing the risk of arbitrary instruction execution from processed content.

nemotron-speech