The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides tools and code examples for serving LLMs that process untrusted user prompts, which creates a potential surface for indirect injection attacks.
Ingestion points: Untrusted data enters the context via the prompts argument in llm.generate() calls (found in SKILL.md and references/optimization.md) and through the /v1/chat/completions API endpoint provided by trtllm-serve (found in SKILL.md and references/serving.md).
Boundary markers: There are no explicit delimiters or boundary instructions (e.g., "ignore embedded instructions") defined in the provided code templates.
Capability inventory: The skill utilizes network capabilities to serve API requests (trtllm-serve) and performs package installation (pip install).
Sanitization: The provided documentation and code snippets do not include input validation or sanitization layers for the prompt data.
[EXTERNAL_DOWNLOADS]: The skill instructs the user to download the tensorrt_llm library from PyPI and the official nvidia/tensorrt_llm Docker image from NVIDIA's container registry. It also references official documentation and repositories on nvidia.github.io and github.com/NVIDIA.
[COMMAND_EXECUTION]: The skill includes instructions for executing shell commands to install dependencies (pip install), pull containers (docker pull), and launch the inference server (trtllm-serve). These are standard operational procedures for the library's intended use case.
[CREDENTIALS_UNSAFE]: The documentation demonstrates secure secret management practices by instructing users to provide API keys via environment variables (e.g., --api_key $API_KEY) rather than hardcoding credentials into configuration files.

tensorrt-llm