Microsoft

Cybersecurity · Artificial Intelligence

Launch

Microsoft launches Project Perception, an agentic cybersecurity platform

Its underlying harness routes tasks to whichever model is cheapest and capable, undercutting rivals' single-model pricing by half.

• Microsoft unveiled Project Perception, an AI cybersecurity system that coordinates red, blue, and green team agents to find, triage, and fix vulnerabilities, entering public preview on August 3, 2026.
• Perception runs on MAI-Cyber-1-Flash, Microsoft's first cybersecurity-specific AI model, paired inside the MDASH bug-hunting harness with OpenAI's GPT-5.4 for the hardest tasks.
• On the CyberGym benchmark, the MAI-Cyber-1-Flash/GPT-5.4 combo scored 96%, beating OpenAI's GPT-5.5 Cyber (85.6%), Anthropic's Mythos 5 (83.8%), and Google's Gemini 3.5 Flash Cyber (83.2%).
• The multi-model harness lets MAI-Cyber-1-Flash handle about 90% of queries on its own and escalates only the toughest 10% to the larger, roughly 10x-bigger GPT-5.4.
• The launch follows recent AI cybersecurity suite debuts from OpenAI and Anthropic, and marks Microsoft's first major security push since EVP Hayete Gallot returned in February to lead the unit.
• As attackers increasingly weaponize AI to exploit flaws at machine speed, Microsoft is betting that cost-efficient orchestration across multiple models, not just raw model power, becomes the deciding edge in AI-driven cyber defense.