Microsoft's new MAI Cyber 2 Flash delivers 92.3% success on CyberGym-E2E, outperforming GPT-6 Astra at a fraction of the cost.

The landscape of AI-driven cybersecurity is shifting from passive scanning to active, agentic defense. On November 17, 2026, Microsoft officially released MAI Cyber 2 Flash, a specialized coding model designed specifically for the rigors of security engineering. Unlike general-purpose LLMs that treat security as a subset of coding, MAI Cyber 2 Flash is purpose-built for vulnerability discovery, automated fix recommendations, and deep-dive threat investigation at a massive codebase scale.
For DevOps and Security engineers, this represents a paradigm shift. We are moving away from simple linting toward autonomous agents capable of reasoning through complex exploit chains. By integrating directly into Microsoft's security ecosystem, this model aims to close the gap between detection and remediation, providing a high-speed, low-cost intelligence layer for modern Security Operations Centers (SOCs).
The most striking architectural upgrade in MAI Cyber 2 Flash is its massive expansion in context handling. The model features a 1M token context window, which is a fourfold increase over its predecessor, MAI Cyber 1.1 Flash. This allows engineers to ingest entire repositories, complex dependency trees, and extensive documentation into a single prompt, enabling the model to understand systemic vulnerabilities that span multiple files.
While Microsoft has not disclosed the exact parameter count or MoE (Mixture of Experts) configuration, the model's efficiency suggests a highly optimized architecture designed for low-latency reasoning. It is specifically tuned to operate within the MDASH framework, Microsoft's multi-model system for advanced threat detection, ensuring that the model's outputs are actionable within existing security workflows.
In head-to-head comparisons, MAI Cyber 2 Flash demonstrates industry-leading efficiency. On the CyberGym-E2E benchmark, the model achieved a 92.3% task success rate, narrowly surpassing GPT-6 Astra's 91.5%. Remarkably, it achieved this result using only half the reasoning budget and at less than 2% of the cost per task. In the WebExploitBench, it scored 40.4%, outperforming GPT-5.6 Cyber's 38.4% under identical reasoning constraints.
To ensure the integrity of these results, Microsoft implemented rigorous evaluation protocols. Out of the 905 tasks in the CyberGym-E2E suite, 92 tasks were identified as overlapping with the training data and were strictly excluded. The reported scores are based on the remaining 813 unique tasks. Furthermore, enterprise customers have reported that the model provides 20% more token efficiency compared to the GPT profile, making it a more sustainable choice for high-volume scanning.
Microsoft has positioned MAI Cyber 2 Flash as an incredibly cost-effective solution for large-scale security automation. The pricing model is highly aggressive, particularly for cached inputs, which is critical for developers running repetitive scans across the same codebases.
Compared to the GPT-6 Astra tier, which costs significantly more for input and output, MAI Cyber 2 Flash provides a massive reduction in TCO (Total Cost of Ownership). This makes it feasible to run continuous, agentic security audits that were previously cost-prohibitive.
MAI Cyber 2 Flash is not a general-purpose chatbot; it is a precision tool. Its primary use case is agentic defense, where it acts as a core reasoning engine for security agents. This includes automated vulnerability discovery in CI/CD pipelines, where the model can flag potential exploits before code is merged.
Beyond discovery, the model excels at 'fix recommendation' workflows. Instead of just identifying a buffer overflow or an injection vulnerability, it can propose a context-aware patch that adheres to the existing codebase's style and safety constraints. It is also ideal for threat investigation, helping SOC analysts traverse large logs and code structures to trace the origin of a breach.
Access to MAI Cyber 2 Flash is currently being rolled out exclusively through Microsoft's security product suite under the codename MDASH. For eligible enterprise customers, the model will begin automatically replacing the current MAI Cyber 1.1 Flash models starting in mid-December 2026.
Developers looking to integrate these capabilities should monitor the Microsoft Security portal and MDASH documentation for upcoming API availability and SDK updates. As the rollout progresses, we expect more direct integration options for third-party security orchestration tools.
API Pricing β Input: 0.10 USD per million tokens / Output: 0.50 USD per million tokens / Context: 1M tokens; 0.01 USD/M for cached input