The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill includes a troubleshooting command in references/troubleshooting.md that requires elevated privileges: sudo ufw allow 8000. This allows the agent to modify system firewall rules.
[REMOTE_CODE_EXECUTION]: The documentation in SKILL.md and references/troubleshooting.md recommends using the --trust-remote-code flag when serving custom models. This vLLM feature allows the execution of arbitrary Python code contained within the model files downloaded from external and potentially untrusted sources like Hugging Face.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external, untrusted data (prompts and API requests) through an LLM engine. It lacks explicit boundary markers or sanitization, creating a surface where malicious instructions in input data could influence agent behavior.
Ingestion points: Prompts are loaded from prompts.txt (SKILL.md) and received via the OpenAI-compatible API endpoint examples (SKILL.md).
Boundary markers: No delimiters or safety instructions are used to separate user prompts from system instructions in the code examples.
Capability inventory: The skill environment allows package installation (pip install), shell command execution (vllm serve), and firewall modification (sudo ufw).
Sanitization: No input validation or filtering is mentioned or implemented in the provided Python snippets.
[COMMAND_EXECUTION]: The skill provides instructions for the agent to execute various shell commands, including environment setup (pip install), model serving (vllm serve), and network diagnostics (curl, lsof, nc).
[EXTERNAL_DOWNLOADS]: The skill instructs the agent to download numerous external packages from PyPI (vllm, locust, autoawq, auto-gptq, flash-attn) and models from Hugging Face. While these target well-known services, they constitute a significant dependency chain.

serving-llms-vllm