localmac-ai-ocr

Warn

Audited by Gen Agent Trust Hub on Mar 7, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes osascript (AppleScript) and ctypes to interface with the macOS ApplicationServices framework for comprehensive GUI automation. This includes activating windows, simulating mouse clicks at specific coordinates, and sending keystrokes to applications. These operations grant the agent significant control over the local desktop environment.
  • [DATA_EXFILTRATION]: The skill captures the local screen using the macOS screencapture utility and transmits the resulting image data (Base64 encoded) to a remote API endpoint defined by the user-provided AISTUDIO_OCR_API_URL environment variable. This constitutes a potential exposure of sensitive information visible on the screen during the capture process.
  • [EXTERNAL_DOWNLOADS]: The setup.sh script and the tool wrappers (scripts/gui, scripts/ocr) use the uv package manager to install and manage Python dependencies such as Pillow and requests. The skill also supports downloading and loading PaddleOCR models if the user chooses the local backend.
  • [INDIRECT_PROMPT_INJECTION]:
  • Ingestion points: The skill processes text extracted from screen captures via OCR in scripts/ocr_tool.py.
  • Boundary markers: No explicit boundary markers or instructions are provided to the agent to treat OCR-extracted text as untrusted content.
  • Capability inventory: The skill possesses high-privilege capabilities including mouse simulation, keystroke injection, and window management in scripts/gui_toolkit.py.
  • Sanitization: Keystrokes processed via send_text are escaped to mitigate basic AppleScript injection, though the logic is primarily for syntax correctness.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 7, 2026, 05:17 PM