Microsoft announced on Monday a new AI security suite called Project Perception and a vulnerability analysis model named MAI-Cyber-1-Flash, which it said achieved a 95.95 percent success rate in CyberGym’s benchmark while reducing costs by roughly half compared with competing commercial models.

According to Microsoft, the MAI-Cyber-1-Flash model, built on the MAI-Thinking-1 reasoning framework and integrated into the MDASH bug‑hunting harness, processes up to 90 percent of security queries internally and hands the remainder to the larger GPT‑5.4 model.

Benchmark results from CyberGym showed MAI-Cyber-1-Flash combined with GPT‑5.4 outperformed OpenAI’s GPT‑5.5 Cyber (85.6 percent), GPT‑5.6 Sol (83.6 percent), Anthropic’s Mythos (83.8 percent) and Google’s Gemini 3.5 Flash (83.2 percent).

Project Perception coordinates three agent categories: red team agents that simulate attack paths, blue team agents that assess risk, and green team agents that remediate issues, aiming to let defenders operate at the speed of attackers.

Microsoft also disclosed the creation of Security FORGE Labs, led by VP of Security Research Taesoo Kim, and an External Red Team Alliance (EXTRA) that will provide unrestricted funding to 18 university research groups worldwide and build a distributed network of specialists for targeted red‑team activities.

The announcements reflect Microsoft’s effort to address growing concerns that AI‑driven threats can bypass traditional security measures, and to give defenders tools that can both detect and patch vulnerabilities automatically.

Microsoft’s AI security initiatives are part of a broader industry push to integrate generative AI into defensive cybersecurity workflows.