Microsoft and Wiz have unveiled rival AI-powered bug-hunting systems that combine multiple large language models to find and fix software vulnerabilities with unusually high accuracy. Both firms say the key to their success is routing different stages of vulnerability analysis to whichever model performs best at that specific task, rather than relying on a single AI system, a lesson with wider implications for how the security industry deploys AI going forward.

Wiz's Project Atlas achieved a 90.9 percent success rate on the CyberGym benchmark and has already uncovered more than 200 zero-day flaws in widely used open-source code, pairing Claude Opus 4.6 with GPT-5.5 and planning to add Gemini next. Microsoft's MDASH harness scored even higher at 95.95 percent by combining its in-house MAI-Cyber-1-Flash model, which handles roughly 90 percent of tasks, with GPT-5.4 for the toughest remaining cases; both beat standalone rivals from OpenAI, Anthropic and Google, which scored between 83 and 86 percent. Microsoft AI chief Mustafa Suleyman said the multi-model approach also roughly halves customer costs compared with using a single frontier model throughout.