The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the mobile applications it automates, which could contain instructions that attempt to influence the agent's behavior.
Ingestion points: The skill uses agent-device snapshot to read accessibility labels, hints, and text properties from the target mobile application's UI.
Boundary markers: The skill does not implement or describe the use of delimiters or boundary markers to isolate UI data from agent instructions.
Capability inventory: The skill possesses significant capabilities including UI interaction (tap, fill), network traffic capture (agent-device network), log retrieval (agent-device logs), and screenshot capture.
Sanitization: No sanitization or filtering logic is provided for data parsed from the mobile app snapshots before it is used by the agent.
[COMMAND_EXECUTION]: The skill's documentation provides code examples where shell commands are constructed using variables derived from untrusted app UI data.
Evidence: In the Node.js/TypeScript example (Example 3), the loginButton.ref property, which originates from an external app snapshot, is directly interpolated into a shell command via execAsync(agent-device tap ${loginButton.ref}) without validation.
Evidence: The skill performs extensive file system writes for artifacts, such as agent-device screenshot ./artifacts/screen.png and agent-device video stop ./artifacts/recording.mp4.
[EXTERNAL_DOWNLOADS]: The skill instructions involve downloading and installing a CLI tool from a public registry.
Evidence: The installation section recommends npm install -g agent-device@latest. The package is hosted on the public npm registry and maintained by a well-known service in the mobile development ecosystem.
[DYNAMIC_EXECUTION]: The skill supports the execution of automation logic stored in external script files.
Evidence: The agent-device replay command allows the agent to execute sequences of instructions stored in .ad files, which may be dynamically created or modified during the agent's lifecycle.

agent-device-mobile-automation