The Agent Skills Directory

[COMMAND_EXECUTION] (MEDIUM): The core functionality of this skill is to send keystrokes to terminal sessions using tmux send-keys. This allows for arbitrary command execution within those sessions. If the agent processes untrusted user input and sends it directly to a shell or REPL, it could lead to command injection.
[REMOTE_CODE_EXECUTION] (MEDIUM): The skill documentation includes examples for remote command execution via ssh. Specifically, it suggests patterns like ssh host 'zsh -c "source ~/.zshrc; ..."', which executes commands on a remote system and sources local configuration files that might contain untrusted or malicious code.
[DATA_EXFILTRATION] (LOW): The use of tmux capture-pane -p in SKILL.md and wait-for-text.sh allows the agent to read the entire output history of a terminal pane. This creates a risk of data exposure if sensitive information, such as passwords, environment variables, or private keys, is displayed in the terminal.
[PROMPT_INJECTION] (LOW): The skill is vulnerable to Indirect Prompt Injection (Category 8):
Ingestion points: Terminal output is read into the agent context via tmux capture-pane in SKILL.md and scripts/wait-for-text.sh.
Boundary markers: None are used to delimit terminal output from agent instructions.
Capability inventory: The skill has high-impact capabilities including tmux send-keys and ssh remote execution.
Sanitization: There is no evidence of sanitization or escaping of the captured terminal text before it is processed by the agent.

tmux