The Agent Skills Directory

[DATA_EXFILTRATION]: The skill demonstrates the use of synchronous HTTP requests (e.g., http.get, http.post) within JavaScript scripts. In the provided examples (e.g., examples/advanced-patterns.md), these are used for legitimate test setup tasks like creating test users on local or designated staging environments (localhost, guideline.io).
[REMOTE_CODE_EXECUTION]: The skill documents the execution of local JavaScript files via the runScript command and inline JavaScript expressions via evalScript or ${} interpolation. This functionality is restricted to the Maestro sandbox (Rhino/GraalJS) and is used to compute dynamic test values or manage test flow logic.
[COMMAND_EXECUTION]: The skill provides an MCP interface for executing Maestro commands and flows (run_flow, run_flow_files) on connected devices. These tools are used for automating mobile application testing and interacting with the device's UI hierarchy.
[PROMPT_INJECTION]: The skill exposes an attack surface for indirect prompt injection by ingesting application data via inspect_view_hierarchy, assertWithAI, and extractTextWithAI. 1. Ingestion points: Screen content is read into the agent context via view hierarchy inspection and AI-powered extraction tools. 2. Boundary markers: None identified in the provided templates; the skill relies on the underlying agent's handling of tool outputs. 3. Capability inventory: The skill includes powerful capabilities such as network requests (http.request), UI interaction (tapOn), and script execution (runScript). 4. Sanitization: There is no explicit sanitization of text extracted from the device screen before it is processed by the agent or used in subsequent steps.

maestro-usage