adithyan-ak/AgentHound

What it solves

AgentHound is an offensive security framework designed to identify vulnerabilities and attack paths within "agentic infrastructure"—the ecosystem of AI agents, model gateways, vector databases, and inference servers. It helps security researchers and red teams discover how a compromised host can be used to pivot into AI services, steal credentials, or execute unauthorized actions through AI tools.

How it works

The framework consists of a standalone collector and an optional analysis server:

1. Collection: A static binary is deployed to a target host. It scans for local agent configurations (e.g., Claude Desktop, Cursor), environment secrets, and reachable AI services (e.g., Ollama, vLLM, Qdrant). It autonomously reuses discovered credentials to verify access to these services.
2. Verification: It performs active validation, such as differential reads on MCP (Model Context Protocol) resources or reversible mutations to prove tool access.
3. Analysis: The collected data is saved as a JSON artifact, which can be ingested into a server that transforms the findings into a queryable attack graph. This graph maps paths for execution, exfiltration, and impersonation.

Who it’s for

It is built for red-team operators and security professionals who are authorized to assess the security of AI-integrated environments and agentic stacks.

Highlights

• Broad Attack Surface: Maps MCP, A2A, model gateways (LiteLLM), inference servers (Ollama/vLLM), vector stores (Qdrant), and MLOps tools (MLflow).
• Credential Reuse: Automatically captures and applies discovered API keys and tokens to unlock further service access during a scan.
• Evidence-Based Proof: Moves beyond reachability guesses by performing actual (and often reversible) reads and invocations to prove access.
• Graph-Native Analysis: Visualizes complex attack paths, such as how a specific credential leads to a tool capable of shell execution or data exfiltration.
• Stealth Mode: Includes a --stealth flag for read-only collection when operational security (OPSEC) is a priority.