scabench-org/hound

What it solves

Hound is an autonomous AI auditor designed to find security vulnerabilities in codebases, specifically optimized for small-to-medium projects like smart contracts. It replaces manual security reviews with an iterative, agentic process that can reason across different subsystems to find complex, cross-component bugs.

How it works

Hound uses a "senior/junior" agent pattern combined with adaptive knowledge graphs. It first builds aspect-oriented graphs (e.g., architecture, value flows) to understand the system. Then, it employs two audit modes: "Sweep Mode" for systematic component analysis and "Intuition Mode" for deep, targeted exploration of high-impact areas like monetary flows.

A "scout" model handles exploration while a "strategist" model manages planning and deep reasoning. The system maintains a belief and hypothesis system where potential vulnerabilities are tracked with confidence scores and evolved through evidence-based investigation. Finally, a reasoning model reviews all hypotheses to confirm or reject them before generating a professional HTML report and PoC prompts.

Who it’s for

Security researchers, smart contract auditors, and developers who need to autonomously scan their code for critical vulnerabilities and generate evidence-backed security reports.

Highlights

• Graph-driven analysis: Builds flexible, agent-designed knowledge graphs to model system architecture and logic.
• Belief & hypothesis system: Tracks observations and assumptions with confidence scores to enable long-horizon reasoning.
• Dynamic model switching: Optimizes cost and performance by using lightweight models for scouting and heavyweight models for strategy.
• Intuition-guided search: Specifically prioritizes high-risk areas like authentication bypasses and state corruption.
• Interactive steering: Includes a chatbot telemetry UI to monitor live activity and steer the agent's investigation in real-time.