The Agent Skills Directory

[COMMAND_EXECUTION]: The pipe-pane --command <shell-command> command allows the agent to execute arbitrary shell commands by piping pane output to them.
[REMOTE_CODE_EXECUTION]: The browser eval <script> and browser addscript <script> commands allow for the execution of arbitrary JavaScript within a browser environment controlled by the agent.
[DATA_EXFILTRATION]: The read-screen and capture-pane commands enable the agent to read the full content of any terminal session, which may contain credentials, private keys, or sensitive configuration data.
[DATA_EXFILTRATION]: Browser control commands such as browser cookies and browser storage provide direct access to sensitive web session data and authentication tokens.
[COMMAND_EXECUTION]: The send command allows the agent to input raw text into any terminal, which can be used to trigger unauthorized command execution if followed by a carriage return.
[PROMPT_INJECTION]: The skill is highly susceptible to Indirect Prompt Injection.
Ingestion points: Terminal content via read-screen and web page content via browser snapshot or browser get.
Boundary markers: No specific boundary markers or sanitization instructions are provided to help the agent distinguish between data and instructions.
Capability inventory: Extensive capabilities including shell command execution (pipe-pane), terminal input (send), and browser script execution (browser eval).
Sanitization: No sanitization or validation of the ingested content is mentioned, allowing external instructions to potentially control the agent's actions.

cmux