The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill promotes secure development by instructing developers to use environment variables for model credentials (ModelConfig.api_key_env) rather than hardcoding secrets. It also explicitly advises redacting credentials, headers, and metadata from serialized configuration mappings before they are emitted to logs.
[PROMPT_INJECTION]: The integration pattern utilizes Fabric().run(input=...) and runtime.invoke(input=...) to process untrusted user data. This creates an indirect prompt injection surface where malicious instructions in the input could attempt to influence the agent harness.
Ingestion points: input and request parameters in SKILL.md used to pass data to agent harnesses.
Boundary markers: The integration guide does not specify boundary markers or delimiters for user-provided input.
Capability inventory: The SDK allows configuring agent harnesses with access to external skills, MCP servers, and models.
Sanitization: Explicit input sanitization or escaping mechanisms are not detailed in the integration documentation, delegating this responsibility to the adapter implementation or the consuming application.
[SAFE]: All installation instructions and repository references, including the use of tools like just and uv, target official and trusted NVIDIA GitHub repositories and documentation sites.

nemo-fabric-integrate