AgentHound
A BloodHound-style risk-analysis tool for multi-agent AI architectures that finds reachable attack paths from untrusted input to sensitive data or external actions.
AgentHound parses multi-agent architecture definitions written for CrewAI, Dify, LangGraph, or a generic YAML authoring format, into a typed capability graph of nodes and edges, then evaluates risk from the combination of capabilities, node properties and missing controls, following a BloodHound-style model. A 35-rule catalog covering path, node and edge patterns classifies each finding as either an attack path or a hygiene and compliance signal.
A FastAPI and NetworkX backend runs analysis in two stages: parsing an uploaded architecture into a stored graph, then running the rule engine to produce findings, risk scores and read endpoints for the graph and findings. A Next.js frontend wizard uploads an architecture file, lets the user review and correct the capabilities the parser inferred, then renders the analyzed graph, findings and risk metrics, including a control-recommendation view that simulates proposed mitigations and shows before-and-after metrics on the mitigated graph.
It ships four progressive example architectures across the three supported frameworks and the generic YAML format, plus a risky and safe scenario pair, as a demo and regression set.