The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes shell commands such as gradlew, curl, and grep to manage the application build and interact with the automation server. This provides the agent with significant control over the local development environment.\n- [COMMAND_EXECUTION]: The library's screenshot endpoint allows the user to specify a file path for the output. If the agent is manipulated into using a sensitive absolute path, it could result in the unauthorized overwriting of local files with image data.\n- [DATA_EXFILTRATION]: The skill enables capturing the application's visual state via screenshots. This allows the agent to access potentially sensitive data displayed in the UI. Furthermore, the embedded server binds to 0.0.0.0 by default, which may expose the control interface to other devices on the same network.\n- [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because the agent reads and interacts with content within the application's UI. If the application displays untrusted data (such as web content or user-submitted text), that data could contain instructions that trick the agent into performing unintended actions using its UI control capabilities.\n- [EXTERNAL_DOWNLOADS]: The skill instructions involve downloading the skill definition from a third-party GitHub repository (forketyfork/compose-ui-test-server) and adding a library dependency from Maven Central.

compose-ui-control