Wiz Unveils Atlas: A Multi-Agent AI System That Discovered Over 200 Vulnerabilities
Wiz has introduced Atlas, an AI-agent-based vulnerability discovery system that has identified over 200 security flaws in open-source projects. The system also secured a $100,000 bounty from GitHub by discovering a critical remote code execution vulnerability (CVE-2026-3854). This development showcases a highly practical application of collaborative multi-agent AI in cybersecurity, proving that specialized AI workflows can outperform single large language models. By scoring 90.9% on the CyberGym benchmark, Atlas surpassed advanced models like GPT-5.5-Cyber and Claude 4.6 in vulnerability research. Atlas utilizes Code Property Graphs (CPGs) to map code structures and employs specialized agents to propose, debate, and adjudicate potential vulnerabilities. To minimize operational costs, it routes routine tasks to smaller models and reserves high-reasoning large models for complex exploit chain analysis and final verification.
## BACKGROUND
A Code Property Graph (CPG) is a graph-based representation of computer programs that merges abstract syntax trees, control flow, and data flow to help identify security flaws. In AI engineering, multi-agent systems divide complex tasks among distinct AI components that collaborate, simulate human workflows, and cross-verify results to reduce errors.