Skip to content

AgentHound Documentation

Attack-path discovery for AI agent infrastructure. BloodHound for MCP, A2A, and AI services.


Get Started

  • Install — Homebrew, Docker, or build from source
  • Quickstart — First scan to first finding in 10 minutes
  • Demo Lab — Full offensive arc with Docker (scan → discover → loot → poison → revert)

Operator Guides

  • Network Scanning — Sweep CIDRs for AI/ML services + fingerprint
  • Protocol Discovery — Find MCP servers and A2A agents by protocol shape
  • Looting — Extract credentials and model artifacts from discovered services
  • LiteLLM — Master key → upstream provider keys
  • Ollama — Model inventory, modelfiles, weight extraction
  • Offensive Actions — Poison tool descriptions, implant configs, revert
  • Attack Paths — Credential chains, cross-protocol pivots, exfiltration routes
  • Deployment — Production setup, reverse proxy, backups
  • Security and OPSEC — Threat model, audit trail, operator posture

Reference

  • CLI Reference — Every command, flag, and env var
  • API Reference — REST endpoints, auth, request/response schemas
  • Graph Model — 25 node types, 25 edge types, ID strategy, merge semantics
  • CAN_REACH — The marquee composite edge (transitive agent→resource access)
  • Detection Rules — 18 pre-built queries + OWASP mapping
  • Rule Syntax — YAML schema for detection + fingerprint rules
  • Configuration — Env vars, state directories, defaults
  • Risk Scoring — Edge weights, node scores, sensitivity classification

Architecture

  • System Design — Two-binary split, data flow, tech stack
  • Ingest Pipeline — Validate → normalize → deduplicate → write → post-process
  • Post-Processors — 11 composite-edge processors with dependency ordering

Contributing

Decisions


Where does my new doc go?

Question Folder
How to USE the tool operationally? operator/
A lookup table, schema, or flag reference? reference/
How the code works internally? architecture/
How to add something to the codebase? contributing/
A first-time setup walkthrough? getting-started/
An architecture decision? adr/

One concept per file. Split before 500 lines. kebab-case filenames.