Microsoft unveils multi-model AI system for vulnerability detection and defense

Microsoft has introduced a group of AI security initiatives that use multiple specialized models and coordinated agents to identify, investigate, and remediate software vulnerabilities.

At the company’s security event, Microsoft presented MDASH, a vulnerability-hunting framework that combines its MAI-Cyber-1-Flash model with the larger GPT-5.4. MAI-Cyber-1-Flash is Microsoft’s first model designed specifically for security analysis and is based on the company’s MAI-Thinking-1 reasoning system.

According to results cited from CyberGym, the combined system achieved a 95.95 percent success rate on the benchmark. The company said this compared with scores of 85.6 percent for OpenAI’s GPT-5.5 Cyber, 83.6 percent for GPT-5.6 Sol, 83.8 percent for Anthropic’s Mythos 5, and 83.2 percent for Google’s Gemini 3.5 Flash Cyber used with CodeMender.

Microsoft said the smaller MAI-Cyber-1-Flash model handles roughly 90 percent of MDASH requests, including finding vulnerabilities, proposing patches, and checking whether fixes are effective. More complex cases are passed to GPT-5.4. The company estimated that this approach reduces operating costs by approximately 50 percent compared with other leading commercial models.

Agent-based defense

Microsoft also announced Project Perception, an agentic security platform organized around three roles. Red-team agents search for weaknesses and model attack paths, blue-team agents assess findings and determine risk, and green-team agents work on remediation.

The company said the system is intended to help defenders respond at a pace closer to that of automated attackers. Details about availability, deployment requirements, and independent validation were not provided in the announcement.

New research and collaboration programs

Microsoft Security FORGE Labs will focus on offensive security research and generative AI. The laboratory will be led by Taesoo Kim, Microsoft’s vice president of security research.

Microsoft also launched the External Red Team Alliance, or EXTRA. Its initial program provides unrestricted funding to 18 university laboratories across six continents for research into AI security and the use of AI in cyber defense. A second component will create a broader network of researchers and practitioners with expertise in particular attack methods, languages, regions, and technical fields.