Behind the AI models built for security at scale | BRK325
Nazmus Sakib and Anthony Penta discuss the engineering choices behind Microsoft’s purpose-built security models, including MAI-Cyber-1-Flash, in this Microsoft Ignite 2026 session. They examine how specialization and evaluation support continuous security analysis without ignoring operational costs or deployment risks.
Building models for continuous security workloads
Always-on defense requires AI that can run continuously. Model capability alone is not enough: accuracy, latency, cost, and trust determine whether a model can be used effectively in security workflows.
The session frames secure AI model development around four pillars:
- Accuracy
- Diversity
- Cost
- Trust
Sakib and Penta connect these considerations to the AI-driven threat landscape and the demands it places on cyber defense. The focus is on specialized models that provide strong security reasoning efficiently, rather than relying solely on general-purpose model capability.
Security specialization and training
Microsoft’s MAI Cyber model strategy combines security data, domain expertise, and reinforcement learning. The presenters cover the development of MAI Cyber One and Cyber 2 Flash models for Microsoft Security, alongside a training pipeline that uses:
- Large-scale repository scans.
- Curated training data.
- Reinforcement learning with verified rewards.
The discussion links this training approach to vulnerability discovery and analysis at scale. It also covers performance results, real-world vulnerability discoveries, and advances in MAI Cyber 2.
Evaluation, safeguards, and controlled access
Rigorous evaluation is central to the session, but the presenters also examine benchmark limitations. Evaluation results are considered alongside the safeguards needed to operate models in enterprise security environments.
Topics include:
- Enterprise security safeguards.
- Curated data practices.
- Controlled model access.
- Guardrails and zero trust containment.
- Safety and governance from training through deployment.
These concerns sit alongside accuracy and efficiency: a model must be suitable for the environment in which it runs, not just capable of performing a security task.
Deployment economics and extensibility
The closing discussion considers cost-efficient deployment and future extensibility. For continuously running security workflows, the trade-offs between accuracy, latency, cost, and trust affect both model selection and practical adoption.
The session presents purpose-built models as an approach to improving AI economics while retaining the security reasoning needed for vulnerability discovery and analysis.
Session chapters
- 00:00 — Introduction to Microsoft Security AI and session goals.
- 00:42 — AI-driven threat landscape and impact on cyber defense.
- 01:36 — Four pillars for secure AI models: accuracy, diversity, cost, and trust.
- 03:26 — MAI Cyber model strategy: security data, expertise, and reinforcement learning.
- 05:11 — Development of MAI Cyber One and Cyber 2 Flash models.
- 06:44 — Enterprise safeguards, curated data practices, and controlled model access.
- 07:53 — Training pipeline: large-scale repository scans and reinforcement learning with verified rewards.
- 10:32 — Benchmark limitations, guardrails, and zero trust containment.
- 14:36 — Performance results, real-world vulnerability discoveries, and MAI Cyber 2 advancements.
- 18:44 — Cost-efficient deployment strategy and future extensibility.