The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to download an installer script for the Android CLI tool from an official domain of a well-known service (dl.google.com). This is documented as a standard setup procedure for the environment.\n- [REMOTE_CODE_EXECUTION]: The installation process involves executing a downloaded shell script. The skill implements security best practices by downloading to a temporary directory, using sed to display the script's content for manual inspection, and requiring explicit user confirmation before execution via bash.\n- [INDIRECT_PROMPT_INJECTION]: The skill includes a feature for evaluating 'journeys' defined in XML files. This creates a vulnerability surface where malicious instructions could be embedded in test data.\n
Ingestion points: The agent reads and processes natural language instructions from external XML files (references/journeys.md).\n
Boundary markers: Instructions are encapsulated in <action> tags, providing structural separation but no protection against the malicious content within the tags. The skill specifically directs the agent to execute these actions 'EXACTLY as written', potentially bypassing internal safety reasoning.\n
Capability inventory: The skill provides access to sensitive operations including adb shell for device interaction, android create for project generation, and android update --url which allows further remote downloads.\n
Sanitization: No input sanitization or validation of the journey action text is described, relying on the agent to interpret potentially adversarial natural language.\n- [COMMAND_EXECUTION]: The skill documents administrative commands for the android tool, including android update --url for self-updating from a user-provided URL and android skills add for installing additional agent skills. While these are documented tool capabilities, they represent powerful functions that allow for the modification of the agent's operating environment.

android-cli