Microsoft Introduces MAI-Cyber-1-Flash and Project Perception
Microsoft announced MAI-Cyber-1-Flash inside its MDASH vulnerability system and introduced Project Perception on July 27, with the agentic security system entering public preview on August 3. Microsoft reports that the new MDASH configuration scored 95.95% on CyberGym and costs nearly 50% less than its current best configuration; those vendor-run comparisons have not been independently replicated.
Microsoft announced MAI-Cyber-1-Flash, its first in-house cybersecurity model, inside the MDASH vulnerability system on July 27. It also introduced Project Perception, a broader agentic security system scheduled to enter public preview on August 3.
The two products are related but distinct. MAI-Cyber-1-Flash is a compact, code-focused model designed to handle most work inside MDASH, Microsoft's multi-model vulnerability identification and remediation harness. Project Perception coordinates security signals, context, models, agents and actions across a wider defensive workflow.
What Microsoft measured
Microsoft reports that an MDASH configuration combining MAI-Cyber-1-Flash with GPT-5.4 scored 95.95% on CyberGym. Its published chart places Mythos at 83.8%, a difference Microsoft rounds to 12 percentage points.
That result is for the combined MDASH system, not MAI-Cyber-1-Flash operating alone. Microsoft says the compact model is designed to handle up to 90% of tasks while the larger GPT-5.4 model is reserved for the hardest 10%. The company also says the combination cuts cost by nearly 50% compared with its current best MDASH setup, which uses GPT-5.4, GPT-5.4 Mini and GPT-5.3 Codex.
These remain vendor-reported benchmark and cost comparisons. Microsoft's public material does not provide enough task-level error analysis or independent replication to treat them as a production performance guarantee.
Red, blue and green agents
Project Perception coordinates three classes of specialized agents. Red-team agents look for paths to compromise, blue-team agents investigate signals and assess meaningful risk, and green-team agents take corrective actions and strengthen defenses.
Microsoft describes those agents as a closed loop supported by context across identities, endpoints, applications, data, clouds and AI systems. The company says the system keeps humans in control and is integrated across Microsoft Security products, but the announcement does not yet provide a complete list of preview workflows or detailed deployment boundaries.
What the preview needs to show
For security and ML teams, the August 3 preview should make the operating details clearer: which workflows are available, how permissions and approvals are configured, what evidence supports each recommendation, and how actions are audited or reversed.
The architecture reflects a shift from AI that summarizes alerts toward systems that can investigate and act. That makes connected context and specialized models more valuable, but also increases the cost of false positives and unsafe actions. Production readiness will depend on permissions, recovery controls and measured action quality as much as on a single benchmark score.
Key Points
- 1Project Perception enters public preview on August 3 and coordinates red, blue and green security agents.
- 2The 95.95% CyberGym result applies to an MDASH configuration combining MAI-Cyber-1-Flash with GPT-5.4, not to the compact model alone.
- 3Microsoft's nearly 50% cost-reduction claim is measured against its current best MDASH configuration and remains a vendor-reported comparison.
Scoring Rationale
Microsoft's first in-house cyber model and the August 3 Project Perception preview could affect AI-assisted vulnerability management and security operations. Availability and architecture are concrete; benchmark, cost and production-effectiveness claims remain vendor-reported.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
