Microsoft on Monday introduced MAI-Cyber-1-Flash, its first in-house cybersecurity model, and Project Perception, an agentic security system that will enter public preview on Aug. 3.
Microsoft said MAI-Cyber-1-Flash is embedded in MDASH, its multi-agent system for finding and fixing software vulnerabilities, and that the setup scored 96% on the CyberGym benchmark while cutting costs by about 50% from its current production configuration.
The company said the benchmark result exceeded frontier models including Mythos, Gemini and GPT. The current MDASH setup uses a mix of GPT-5.4, GPT-5.4 mini and GPT-5.3 codex.
Project Perception coordinates red team agents that search for paths to compromise, blue team agents that investigate and triage risk, and green team agents that remediate and harden defenses, Microsoft said.
In an interview with VentureBeat, Microsoft AI CEO Mustafa Suleyman said the announcement was the start of a broader push. “We really do have a pretty significant data and harness and expertise moat,” he told VentureBeat.
Suleyman said MAI-Cyber-1-Flash was built to handle up to 90% of security tasks, while MDASH sends the remaining 10% of harder problems to OpenAI’s GPT-5.4. He described the harness as a router that matches incoming problems to the model best suited to solve them.
Asked why Microsoft uses GPT-5.4 for the escalation tier, Suleyman cited cost. “GPT-5.6 is expensive. GPT-5.4 is incredibly good relative to its cost,” he told VentureBeat.
He said enterprise customers were pushing back on frontier-model pricing as token costs rise. “The top model providers want you to use the most expensive model continuously, whereas because we are a platform, we’re on the side of the enterprise,” Suleyman said.
Microsoft said it processes more than 100 trillion security signals a day and draws on data from 1.6 million customers. Its 2025 Digital Defense Report cited 4.5 million new malware files blocked and 5 billion emails screened per day.
Suleyman said that telemetry gives Microsoft a competitive advantage. “That is definitely a moat for us,” he said, referring to the company’s data, expertise and operating experience.
Microsoft said access to the vulnerability-hunting model will be tightly controlled because the system could be misused by attackers. Suleyman said the company will monitor API use and roll out access in stages, starting with tens of users, then hundreds, then thousands.
The company said the model was evaluated by Microsoft’s AI Red Team, tested through automated and expert-led adversarial exercises, and assessed by an independent third party. Microsoft said deployment includes tenant isolation, auditing and sandboxed execution environments with no internet access.
Suleyman also said he was skeptical that enterprise AI would converge on a single large model. “It remains to be seen whether one giant model that is fully multimodal is actually able to deliver additional transfer learning benefit because of the integration, or whether it’s just a big lumbering expensive giant,” he told VentureBeat.
