The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes a workflow where the agent ingests data from external CLI processes, which constitutes a vulnerability surface for instructions embedded in tool output.
Ingestion points: The agent is instructed to use tmux capture-pane -pt and os.read(master_fd, 4096) in SKILL.md to monitor the state of the terminal.
Boundary markers: The provided templates do not include explicit delimiters or "ignore embedded instructions" markers when reading terminal buffers.
Capability inventory: The skill uses subprocess.Popen, tmux send-keys, and os.write to interact with the system, providing a mechanism for an injected instruction to trigger follow-up actions.
Sanitization: The skill includes a "Guardrails" section that explicitly warns against sending credentials or destructive commands, which acts as a manual safety check but does not programmatically sanitize the input.
[DYNAMIC_EXECUTION]: The skill provides templates for generating and executing temporary Python and Bash scripts at runtime to manage PTY (Pseudo-Terminal) sessions.
Evidence: SKILL.md contains code blocks for a "Minimal PTY Harness" using Python's pty and subprocess modules, and a "Minimal tmux Harness" using shell scripting.
Severity Context: This is classified as simple script generation from known templates for legitimate automation purposes, but it involves the runtime creation of executable logic.

cursor-control-cli