doc-inject-guard
CLI tool that scans documents (DOCX, PDF, Markdown, HTML) for embedded prompt injection payloads targeting AI agents. Features five detection modules, risk scoring, multi-format reporting, and watch mode.
doc-inject-guard
Scan documents for hidden prompt injection before they reach your AI agent pipeline.
The Context Collapse AI worm (2026) proved malicious instructions hidden in Word documents can alter financial data via Microsoft Copilot, exfiltrate sensitive information, and self-propagate. doc-inject-guard acts as a document-level security gate.
Quick Start
pip install doc-inject-guard
doc-inject-guard scan ./incoming-docs/
doc-inject-guard scan ./incoming/ --recursive --format json --ci
doc-inject-guard analyze suspicious.docx --llm
doc-inject-guard watch ./hotfolder/
Features
- 4 parsers — DOCX, PDF, Markdown, HTML
- 5 detection modules — Hidden text, encoded payloads, suspicious URLs, metadata injection, behavioral instructions
- Risk scoring — 0-100 weighted score with severity categorization
- 3 output formats — Rich terminal, JSON (CI), SARIF (GitHub Security)
- Watch mode — Real-time directory monitoring
- LLM analyzer — Optional deep analysis for subtle injection patterns
- 112 tests, 85% coverage