android-use
Pass
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill reads and processes accessibility XML from Android devices using the
get-screen.shscript, creating a surface for injection from untrusted app content. - Ingestion points: UI hierarchy XML containing attributes like
textandcontent-desc(SKILL.md). - Boundary markers: Absent; the instructions direct the agent to "Read the accessibility tree to understand what's on screen" without delimiters or instructions to ignore embedded commands.
- Capability inventory: Significant capabilities including executing bash scripts for screen interaction, keystroke injection, and application management (SKILL.md).
- Sanitization: Absent; the agent is expected to decide actions based on raw XML content which may contain malicious instructions designed to look like UI elements.
- [COMMAND_EXECUTION]: The skill relies on executing a suite of local bash scripts (
scripts/tap.sh,scripts/type-text.sh,scripts/launch-app.sh, etc.) to interact with the host system and connected Android hardware via ADB. - [EXTERNAL_DOWNLOADS]: Downloads the skill definition from the author's GitHub repository during the installation process.
Audit Metadata