Skip to main content
IPI tests whether documents ingested by AI systems can redirect agent behavior through hidden instructions embedded in the document content.

​
Why IPI Matters

AI pipelines that ingest external documents — RAG systems, knowledge bases, web fetch tools — are vulnerable to adversarial content in those documents. Unlike output analysis, IPI confirms execution via out-of-band HTTP callbacks: the payload fires a request to your listener, providing proof of execution independent of model output.

​
How It Works

The IPI workflow follows five steps:
  1. Start the listener — Launch the callback server on your machine to receive execution confirmations
  2. Generate payloads — Create payload documents using a technique and format combination
  3. Deploy — Place the document in the target pipeline (upload to knowledge base, RAG corpus, or web-accessible URL)
  4. Wait for callbacks — The callback fires when the agent ingests and acts on the payload
  5. Review results — Check execution status in the web dashboard or via ipi status

​
Built-in Components

  • 34 techniques across 3 categories: social engineering, instruction override, context manipulation
  • 7 output formats — Markdown, plain text, HTML, PDF, DOCX, CSV, JSON
  • Callback server — Authenticated listener with HMAC verification
  • Web dashboard — Real-time callback monitoring and campaign management
  • Deterministic seeding — Reproducible payload generation for consistent testing

​
Next Steps

  • IPI CLI Reference — Command reference for countersignal ipi
  • Techniques — Social engineering, instruction override, and context manipulation techniques
  • Formats — Supported output formats for payload generation
  • Payloads — Payload structure and customization
  • Callbacks — Callback server setup and HMAC verification
  • Web Dashboard — Real-time monitoring interface