box-legal-workflows-ma

Pass

Audited by Gen Agent Trust Hub on Sep 7, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill documentation provides instructions to fetch and install additional modules from the author's official GitHub repository (https://github.com/box/box-for-ai). These references are part of the intended modular setup and target the vendor's own infrastructure.
  • [REMOTE_CODE_EXECUTION]: Prerequisite steps involve using a CLI tool to add skills directly from a remote GitHub URL. While this involves executing remote content, the source is the author's official repository, which is a recognized service for this skill's ecosystem.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze and extract data from external submissions and due-diligence documents, creating a surface for indirect prompt injection.
  • Ingestion points: External documents and submissions are processed via ai_qa_multi_file, ai_qa_single_file, and ai_extract_structured_from_fields_enhanced (SKILL.md).
  • Boundary markers: While the skill references external boundaries in box-legal-workflows, it does not explicitly define unique delimiters or "ignore instructions" tags within the prompt templates provided in this file.
  • Capability inventory: The skill can perform file and folder operations (create_folder, upload_file, copy_file), modify access controls (create_collaboration, add_folder_shared_link), and update file metadata (set_file_metadata) (SKILL.md).
  • Sanitization: The workflow mitigates risk by enforcing manual human confirmation steps for folder hierarchy, role assignments, and external sharing actions.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 7, 2026, 06:17 AM
Security Audit — agent-trust-hub — box-legal-workflows-ma