Microsoft AI cybersecurity model targets enterprise security costs

Microsoft announced its first custom Microsoft AI cybersecurity model Monday, pairing MAI-Cyber-1-Flash with an agentic defense system meant to lower the cost of enterprise vulnerability work. The company says the model, built by Microsoft AI, sits inside MDASH, its multi-agent system for finding and fixing software flaws, and helped the configuration score 96% on CyberGym while cutting costs by roughly half versus Microsoft’s current production setup.

Microsoft also introduced Project Perception, a public-preview system scheduled for August 3 that coordinates three types of agents: red-team agents that look for compromise paths, blue-team agents that investigate and triage risk, and green-team agents that apply fixes and harden defenses.

How does Microsoft’s AI cybersecurity model work?

MAI-Cyber-1-Flash is a compact security model designed to take on most routine security reasoning inside MDASH. Microsoft says the system routes about 90% of tasks to the smaller model and sends the remaining, more difficult 10% to OpenAI’s GPT-5.4, which remains part of the stack for the hardest coding work.

In an interview with VentureBeat, Microsoft AI CEO Mustafa Suleyman described MDASH as the routing layer that decides which model should handle each problem. He said the system combines the harness, the faster Microsoft model and GPT-5.4, which he called a generalist coding model. His argument is that the system’s performance comes from long agentic workflows involving state, external data, code generation and validation, rather than from a single model outperforming every rival model on its own.

That architecture is also a cost statement. Suleyman told VentureBeat that GPT-5.6 is expensive and that GPT-5.4 offers strong capability relative to its cost. Microsoft says the new MDASH configuration is about 50% cheaper than its current setup, which uses a mix of GPT-5.4, 5.4 mini and 5.3 codex. The company framed the issue as an enterprise buying problem: security workloads run continuously, and token costs grow quickly when every task is sent to the most expensive model available.

Microsoft says the MDASH configuration reached 96% on CyberGym, a benchmark for testing how AI systems reason across large codebases to find real vulnerabilities. The company says that result topped systems based on Mythos, Gemini and GPT. The comparison needs context: the result comes from Microsoft’s own evaluation, and the measured system is a tuned harness plus multiple models, rather than a clean model-to-model benchmark.

Microsoft is leaning on a data advantage it says other labs cannot easily match. The company says it processes more than 100 trillion security signals a day and draws operational insight from 1.6 million customers. Its 2025 Digital Defense Report also cited 4.5 million new malware files blocked and 5 billion emails screened daily. Suleyman told VentureBeat that Microsoft has decades of cybersecurity data and institutional experience, including from customers in government that have faced sustained attacks.

The product also carries an obvious dual-use risk. A system that can find vulnerabilities in complex software can help defenders, and the same capability can interest attackers. Microsoft says access will be gated, API use will be monitored, and rollout will start with small numbers of approved users before expanding. The company also says the model was reviewed by its AI Red Team, tested through automated and expert-led adversarial exercises, assessed by a third party, and deployed with tenant isolation, auditing and sandboxed execution environments without internet access.

For Microsoft, the announcement points to a broader AI strategy: specialized models, orchestration and proprietary telemetry, with frontier models used selectively when the work justifies the cost. That approach is strongest in security, where Microsoft already controls products that observe threats and tools that can act on them. The open question is how well the same playbook works in categories where its data advantage is thinner.

This story draws on original reporting from VentureBeat.