audit-report-checker

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.30). Skill “audit-report-checker” at runtime ingests the user-provided audit report content (Step 1 parse → report.md/extracted_tables.json → Step 2 reads report.md with Claude and Step 4/5 sends extracted table text and original text snippets to the DeepSeek/Claude AI for semantic localization and review), so outsider-authored free text inside the document is LLM-readable and can carry indirect prompt-injection payloads.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The skill makes runtime calls to external AI and OCR services — DeepSeek at "https://api.deepseek.com" (used by scripts/ai_worker.py to send prompts and receive JSON annotations) and Mineru at "https://mineru.net" / "https://mineru.net/apiManage/token" (invoked via the mineru-open-api subprocess to upload documents and return extracted markdown/OCR), so remote content directly influences agent behavior and parsing.