The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill uses sudo to perform high-privilege system modifications, including disabling system sleep (pmset), enabling Remote Login (systemsetup), configuring the macOS Application Firewall (socketfilterfw), and increasing system-wide file descriptor limits (launchctl limit).
[PERSISTENCE]: The skill establishes persistence by creating a launchd agent (com.ollama.serve.plist) in the user's LaunchAgents directory to ensure the LLM service starts at login and remains active. It also instructs the user to add a health check script to crontab for periodic execution.
[COMMAND_EXECUTION]: The skill modifies sensitive system configurations, specifically the SSH daemon configuration (/etc/ssh/sshd_config), and creates executable shell scripts (llm-health.sh) containing monitoring logic.
[DYNAMIC_EXECUTION]: The skill includes an inline Python script executed via python3 -c to demonstrate model loading and generation using the MLX framework. Additionally, it contains a command (curl ... | python3 -m json.tool) that pipes output from a local API to the Python JSON module for formatting.
[EXTERNAL_DOWNLOADS]: The skill fetches and installs software from external sources, including Homebrew packages, Tailscale for remote access, and Docker images from the GitHub Container Registry (ghcr.io).
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface where it reads and processes data from a local API (http://localhost:11434/api/ps) and passes it to the Python interpreter for formatting. The health check script also executes system commands based on the availability of this local service, representing a potential feedback loop for injected data if the local API is compromised.

mac-mini-llm-lab