red-flag-contract-scanner

Warn

Audited by Gen Agent Trust Hub on May 5, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill generates and executes Python scripts at runtime to produce Word-formatted reports. As described in assets/docx_template_example.md, the agent is instructed to use the python-docx library to assemble a script that includes metadata and specific clause text extracted from user-supplied documents. This dynamic assembly and execution of code based on external inputs is a medium-risk pattern.\n- [PROMPT_INJECTION]: The skill is designed to ingest and interpret large amounts of untrusted external text (legal contracts, terms of service, and privacy policies). This establishes a significant surface for indirect prompt injection.\n
  • Ingestion points: Untrusted contract text is ingested when a user pastes or uploads a document for analysis (e.g., in Step 1 of the workflow in SKILL.md).\n
  • Boundary markers: The instructions do not define strict delimiters or 'ignore embedded instructions' warnings to separate the contract data from the agent's internal analysis logic.\n
  • Capability inventory: The agent has the capability to execute generated Python code (via the Word report generation process) and compose draft emails (via the message_compose_v1 tool mentioned in SKILL.md).\n
  • Sanitization: There is no evidence of validation or sanitization of the contract text before it is used for report generation or decision-making.
Audit Metadata
Risk Level
MEDIUM
Analyzed
May 5, 2026, 08:57 AM