The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's primary purpose is to facilitate command execution and interaction within remote web terminals. It outlines several methods (textbox input, keyboard simulation, and direct JavaScript evaluation) to send commands to target environments such as Cloud Shells or CTF ranges.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from potentially untrusted external sources (the output of a remote web terminal), which introduces a vulnerability surface where the agent could receive instructions from the target environment.
Ingestion points: Terminal content is retrieved from the browser DOM (.xterm-rows, .xterm-screen) or the terminal's internal memory buffer via browser_evaluate as described in Phase 3 of SKILL.md.
Boundary markers: The instructions explicitly recommend a "marker wrapping" technique—using specific strings like ===START=== and ===END=== around command outputs—to help the agent differentiate between intended results and unrelated terminal noise or potential injection attempts.
Capability inventory: The skill leverages the agent's browser automation toolset, including browser_type, browser_press_key, browser_evaluate, and browser_take_screenshot.
Sanitization: The skill focuses on detection and extraction accuracy; it relies on the suggested marker strategy rather than automated sanitization of the terminal stream.
[DATA_EXPOSURE]: The skill includes examples of setting environment variables and accessing configuration files (e.g., AWS credentials) in the context of cloud shell administration. All examples use standard non-functional placeholders (like AKIAXXXX), adhering to safe documentation practices.

browser-xterm-interaction