visa-doc-translate
Warn
Audited by Gen Agent Trust Hub on Oct 14, 2026
Risk Level: MEDIUMDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill instructs the agent to dynamically create and execute a Python script using the
PILandreportlablibraries to generate a bilingual PDF. The content of this script is derived from the OCR and translation of the input document. If a document contains malicious text designed to look like code or instructions, it could potentially influence the generated script, leading to arbitrary code execution on the user's machine. - [EXTERNAL_DOWNLOADS]: The skill requires the installation of multiple third-party Python packages, including
pillow,reportlab,easyocr,pytesseract,pyobjc-framework-Vision, andpyobjc-framework-Quartz. It also suggests installing system-level dependencies liketesseractvia Homebrew. While these are common libraries, the automatic installation and reliance on external code increase the attack surface. - [COMMAND_EXECUTION]: The skill executes the
sipscommand-line utility to convert HEIC images to PNG format. It also executes shell commands to install dependencies if they are missing. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input in the form of document images (bank statements, passports, etc.). The text extracted via OCR is directly used in the translation prompt and the subsequent PDF generation script. There are no boundary markers or sanitization steps mentioned to prevent 'jailbreak' or 'injection' text within a scanned document from overriding the agent's instructions.
- [DATA_EXPOSURE]: The skill is designed to handle highly sensitive personal information, including bank certificates, ID cards, and passports. While no explicit network exfiltration was found, the processing of this data through an LLM and local scripts without strict data handling policies represents a privacy risk.
Audit Metadata