Google Wiz launches Atlas AI agent to find bugs, already found over 200

Google's cybersecurity division Wiz has launched Atlas, an AI agent system for finding security bugs. It uses multiple AI agents working together like a human security team, and has already found over 200 vulnerabilities. Atlas breaks down tasks into stages. It first builds a threat model for the target code, then uses code property graphs to find potential attack paths. Different AI agents propose and check each other's bug hypotheses. Only bugs that pass review by three roles—supporting argument, counter-analysis, and final decision—are kept. To save costs, Atlas uses bigger AI models for complex tasks and smaller ones for routine work. It also uses deterministic code analysis and automatically sets up test environments to verify bugs. In the CyberGym benchmark, Atlas scored 90.9%. In real use, Atlas has found over 200 unknown bugs in major open-source projects. It also helped GitHub find a serious remote code execution bug (CVE-2026-3854), earning a $100,000 bounty.