Microsoft has introduced MAI-Cyber-1-Flash, its first artificial intelligence model designed specifically for cybersecurity tasks. The model runs exclusively inside MDASH, a multi-model vulnerability identification and remediation system. According to Microsoft’s announcement, the MDASH configuration with MAI-Cyber-1-Flash and GPT-5.4 scored 95.95% on the CyberGym benchmark, while cutting system operating costs by 50% compared to the previous model combination. Access is limited to approved MDASH customers through a private preview in Azure AI Foundry.
Architecture and routing approach
According to the model card, MAI-Cyber-1-Flash is a sparse transformer based on a Mixture-of-Experts (MoE) architecture with 137 billion parameters, of which only 5 billion are active at any given time. The context window is 256,000 tokens. The model is a cybersecurity-specialized refinement of MAI-Code-1-Flash, which in turn was created from an intermediate MAI-Thinking-1 checkpoint.
The key technical idea is task routing. MAI-Cyber-1-Flash is intended to handle up to 90% of MDASH tasks, while GPT-5.4 is invoked only for the most complex 10%. According to the model card, the new configuration has replaced 80% of the models previously used in the system. As Taesu Kim, Microsoft’s vice president for agent security, wrote when describing MDASH in June: “The model is one of the inputs; the system around it is the product.”
What the benchmarks actually measure
The reported 95.95% is the result of the MDASH system with MAI-Cyber-1-Flash and GPT-5.4 combined, not a standalone metric for the new model. This is a fundamental distinction that is easy to overlook.
CyberGym Level 1 is a test of reproducing known vulnerabilities: the agent is given a vulnerability description and unpatched source code, and then evaluated on whether it can create a working proof-of-concept. The benchmark does not measure the ability to discover unknown vulnerabilities and does not validate the correctness of generated patches.
Several factors call for extra caution when interpreting the results:
- Missing from the public leaderboard. As of July 28, 2026, the public CyberGym leaderboard still showed Microsoft’s previous score — 88.4% from May 12. It is not specified whether the new score was ever submitted for listing.
- Not comparable with the June result. Microsoft previously reported a 96.55% score for MDASH, but that result counted any crash, including non-target vulnerabilities. The July materials do not disclose whether the same criterion was applied to the 95.95% figure, so treating the two numbers as a performance trend is incorrect.
- Opaque cost savings. The claimed 50% cost reduction is defined relative to the prior best MDASH configuration (GPT-5.4, GPT-5.4 mini, GPT-5.3 Codex). However, Microsoft did not disclose token consumption, call volume, latency, task distribution, or compute resources, making it impossible to reproduce or normalize this figure.
Results beyond CyberGym
The model card includes MAI-Cyber-1-Flash results under a simplified terminal shell (without the full MDASH system), which present a more restrained picture:
- CVEBench — 0.314
- CyberSecEval4 (threat analysis) — 0.553
- Malware analysis test — 0.33
- CRSBench (POV=1200) — 0.651
- ExploitGym (kernel, user space, browser) — zero across all categories
ExploitGym requires the agent to turn known vulnerabilities and crashing inputs into working code execution exploits. Zero scores in all categories indicate that the model is currently unable to solve exploit development tasks — a significant limitation that should be taken into account when assessing the overall maturity of the technology.
Microsoft notes that all testing was conducted in a network-isolated environment with no access to production systems, the internet, or external services. The model card also includes a warning: generated text and code may be inaccurate or incomplete and must be reviewed before use in critical scenarios.
Place in Microsoft’s strategy: Project Perception
Vulnerability management via MAI-Cyber-1-Flash inside MDASH is the first announced scenario for Project Perception, a broader system for coordinating defensive security agents. According to Microsoft’s announcement, a public preview of Project Perception is scheduled for August 3, followed by extending the model to additional security workflows beyond vulnerability management.
Practical takeaways
For organizations already using MDASH, the new configuration is potentially attractive due to lower costs at comparable quality levels — but until independent confirmations appear, the reported numbers remain vendor claims. For the rest of the market, MAI-Cyber-1-Flash is not yet available: the model is not offered as a standalone product or via a public API.
The main takeaway from this release is not the specific percentages, but the architectural approach: Microsoft is betting on routing between a specialized compact model and a powerful general-purpose one, optimizing cost–quality trade-offs at the system level. At the same time, the zero scores on ExploitGym and moderate performance on other benchmarks without the MDASH shell highlight that the value is created by the system, not the model by itself. Organizations evaluating AI tools for vulnerability management should ask vendors not only for headline benchmarks, but also for their evaluation methodology, testing conditions, and results on tasks that are close to real-world exploitation scenarios.
