Microsoft Project Perception: The AI Security Tool That Found 570 Vulnerabilities in One Month

Microsoft is preparing to release Project Perception, a commercial AI-powered cybersecurity tool that scans enterprise systems, identifies software vulnerabilities, and automatically generates fixes using a combination of AI models from Anthropic (Claude), OpenAI (GPT-5.6), and Microsoft's own models . First reported by The Information on July 16, 2026, the tool is expected to launch by the end of July 2026 .

How Project Perception Works

Unlike single-model security tools, Project Perception uses a "model router" — a dynamic selection system that chooses the cheapest or most appropriate AI model for each subtask. This multi-model architecture optimizes both cost and performance, making it a cost-effective alternative to competing products like Anthropic's Mythos .

Caveat: Microsoft has not yet made a public announcement, published technical documentation, or confirmed pricing, supported environments, or data-handling terms. All details currently come from press reports .

New Leadership Driving the Push

Project Perception arrives under new security leadership. In February 2026, Microsoft CEO Satya Nadella announced that Hayete Gallot would rejoin Microsoft as Executive Vice President of Security, replacing Charlie Bell, who moved to a new engineering quality role reporting directly to Nadella . Since taking over, Gallot has replaced several senior security executives as part of a broader AI restructuring within the division, according to The Information .

The Broader AI Security Ecosystem

Project Perception is one part of a larger AI-driven cybersecurity strategy at Microsoft:

Component	Description
MDASH (Multi-Model Agentic Scanning Harness)	A multi-model agentic scanning system that orchestrates specialized AI agents to discover, validate, and remediate software vulnerabilities. Put into public preview in June 2026 .
Agentic security system (May 2026)	An earlier agentic system that helped researchers find 16 new vulnerabilities across Windows networking and authentication, including 4 Critical RCE flaws .
July 2026 Patch Tuesday	Microsoft released a record 570 patches in a single Patch Tuesday cycle — a massive jump driven in part by the automated vulnerability discovery capabilities of MDASH and related AI tools .

Key Facts at a Glance

• Product: Project Perception — a commercial AI vulnerability scanner
• Models used: Anthropic (Claude), OpenAI (GPT-5.6), Microsoft's own models
• Key feature: Cost-optimizing model router that dynamically selects the best AI for each task
• Launch window: End of July 2026 (per press reports)
• Security lead: Hayete Gallot, EVP of Security since February 2026
• Related system: MDASH — multi-model scanning harness already in public preview
• Record patch cycle: ~570 patches in July 2026 Patch Tuesday, linked to AI-driven vulnerability hunting

What's Still Unknown

While press reporting is consistent and well-sourced, Microsoft has not yet officially confirmed Project Perception, MDASH's commercial roadmap, pricing, or enterprise data-handling terms. All current details originate from The Information and corroborating outlets like TechRepublic .