AgentHound
An offensive security framework for agentic infrastructure that captures credentials and configuration from a compromised host, fingerprints reachable AI services such as MCP servers and model gateways, and builds an attack graph.
AgentHound drops one static collector onto a host and runs a single scan that gathers local credentials and agent configuration, discovers reachable AI services across MCP, A2A, model gateways, inference servers, vector stores, MLOps tools and notebooks, fingerprints and inventories them, reuses compatible credentials, and verifies concrete access, checkpointing everything into one JSON artifact. It targets systems the user owns or is authorized to assess.
The default scan is active, since a foothold may disappear at any time, while a stealth mode switches to read-only collection when needed. For an eligible MCP resource, it first attempts an anonymous read before trying credentialed access, so it records verified evidence rather than guessing at reachability. An optional analysis server turns the collected artifact into a queryable attack graph with credential chains, execution and exfiltration paths, cross-protocol pivots, risk scoring and history, and it can perform reversible active validation against eligible managed tools.