- MAI-Cyber-1-Flash is a sparse mixture-of-experts model with 137 billion total parameters, 5 billion active parameters and a 256,000-token context window. [1] - Microsoft says the model can handle up to 90% of MDASH tasks, with the hardest 10% routed to OpenAI’s GPT-5.4. [2] - The reported 95.95% CyberGym result measures MDASH with multiple models and tools, not MAI-Cyber-1-Flash alone. As of July 28, CyberGym’s public leaderboard still showed Microsoft’s older 88.4% result. [1][5] - Microsoft claims the configuration costs 50% less than MDASH’s previous GPT-based model mix, but it did not disclose prices or enough workload data to reproduce the calculation.
[2] Microsoft launched MAI-Cyber-1-Flash on July 27 as its first model developed specifically for cybersecurity workflows. The company designed it to operate inside MDASH, a multi-model system that uses specialized agents to discover, validate and help remediate software vulnerabilities.[1][2]
The model reduces rather than removes MDASH’s reliance on frontier models. Microsoft says MAI-Cyber-1-Flash can process up to 90% of tasks, while the 10% deemed exceptionally difficult are sent to OpenAI’s GPT-5.4 in the configuration used for the launch benchmark.[2]
The 96% result belongs to the system #
Microsoft says MDASH equipped with MAI-Cyber-1-Flash and GPT-5.4 achieved a 95.95% success rate on CyberGym. The model card says replacing 80% of MDASH’s previous model mix with MAI-Cyber-1-Flash raised the system’s score from 88.4% to 95.95%. Microsoft separately says the model is designed to handle up to 90% of MDASH tasks. The figures describe different measures—the share of models replaced in the evaluation and the expected share of tasks handled by the smaller model—but the launch materials do not provide the underlying routing distribution.[1][2]
That 95.95% result was not reflected on CyberGym’s public leaderboard as of July 28. The benchmark site still listed an independent system at 89.6% and Microsoft’s previous MDASH configuration at 88.4%. The new result currently appears in Microsoft’s launch materials and model card rather than as a separately posted benchmark entry.[1][5]
CyberGym contains 1,507 historical vulnerabilities from 188 software projects. Its principal task gives an agent a vulnerability description and the corresponding unpatched codebase, then asks it to generate a proof-of-concept input that reproduces the flaw. Microsoft used CyberGym’s level-one configuration, which provides vulnerable source code and a high-level description.[3][4]
The benchmark requires codebase-level reasoning and working proof-of-concept generation, but it is primarily a targeted vulnerability-reproduction test. A 95.95% score should not be read as a 95.95% success rate for discovering previously unknown vulnerabilities without clues.[3][4]
Microsoft’s model card also reports standalone results from a lightweight terminal harness: 0.314 on CVEBench, 0.651 on CRSBench, 0.553 on CyberSecEval4 threat intelligence, 0.33 on its malware-analysis portion and zero in the kernel, userspace and browser categories of ExploitGym. Microsoft did not publish competitor results under the same standalone setup. Those figures are not directly comparable with the MDASH CyberGym result, which benefits from more than 100 agents, multiple models and a specialized orchestration pipeline.[1][3]
The cost claim uses Microsoft’s previous model mix as its baseline #
Microsoft says MAI-Cyber-1-Flash cuts model costs by 50% compared with its previous best MDASH configuration, which used GPT-5.4, GPT-5.4 mini and GPT-5.3 Codex. The comparison is narrower than the product page’s broader description of delivering comparable performance at half the cost of “leading models.”[2][8]
The company did not disclose token prices, average cost per repository, latency, the number of model calls or the precise proportion of work routed to each model. The claimed savings therefore cannot be independently reproduced from the launch documents. Customer savings could also vary with codebase size, task difficulty and how frequently MDASH escalates work to GPT-5.4.[1][2]
MAI-Cyber-1-Flash itself is a cybersecurity fine-tune of MAI-Code-1-Flash. It uses sparse mixture-of-experts layers, with 137 billion total parameters but 5 billion active during inference, and supports a 256,000-token context window. Microsoft’s launch post also describes it as part of the broader MAI-Thinking-1 development lineage.[1][2]
MDASH contributes much of the engineering #
MDASH processes a repository through prepare, scan, validation, deduplication and proof stages. Specialized agents analyze attack surfaces, propose candidate flaws, debate whether findings are reachable and exploitable, group duplicates, and attempt to construct inputs that trigger the suspected vulnerability. Microsoft’s description makes clear that the model is one component of the product rather than the complete vulnerability scanner.[3]
Microsoft introduced MDASH in May after using it in research that contributed to 16 Windows CVEs included in Microsoft’s May 12 security updates, including four critical remote-code-execution vulnerabilities. The company also reported finding all 21 planted vulnerabilities in a private test driver with no false positives in that run. The planted-driver evaluation and historical recall tests were designed and conducted by Microsoft, while the 16 reported findings received CVE identifiers and were shipped through Microsoft’s update process.[3]
Microsoft says MDASH is used by its own security engineering teams and was tested by a small set of customers in a limited private preview. Its current product page goes further, saying MDASH is deployed by leading Fortune 500 companies, but it does not name them or describe their deployments.[3][8]
Microsoft has named Nationwide Building Society as a Project Perception user. That customer account concerns the broader red-, blue- and green-agent security platform, however, and does not establish that Nationwide used MAI-Cyber-1-Flash for production vulnerability scanning.[6]
Access remains restricted #
The model card says MAI-Cyber-1-Flash is available through an Azure AI Foundry private preview solely for use within MDASH. Access is restricted to selected customers and requires additional approval because vulnerability discovery and exploit validation can be used by attackers as well as defenders. Uses outside defensive operations in MDASH are designated out of scope.[1]
Microsoft says the model was calibrated to reject ambiguous or potentially harmful requests, which may also cause it to decline some legitimate defensive work. The model card warns that generated text and code can be inaccurate or incomplete and should be reviewed, tested and validated before consequential use.[1]
MDASH adds role-based controls, tenant isolation, encryption, auditing and sandboxed execution without internet access. Microsoft also says its red team tested the model in isolated environments using a system drawing on more than one million multi-turn jailbreak conversations. An independent third party found no critical-severity issues, according to the model card, but Microsoft did not name the assessor or publish its methodology or findings.[1][2]
Project Perception, Microsoft’s broader agentic-security platform, is scheduled to enter public preview on August 3, 2026. Microsoft says it will eventually use MAI-Cyber-1-Flash for security workflows beyond software vulnerability analysis, but the model card currently limits the model’s distribution to approved MDASH use.[1][7]
Companies mentioned #
Further sources #
[[1] Microsoft, “MAI-Cyber-1-Flash Model Card,” July 27, 2026. ↗](https://microsoft.ai/pdf/MAI-Cyber-1-Flash-Model-Card.pdf)
[[2] Microsoft AI, “Introducing MAI-Cyber-1-Flash inside MDASH,” July 27, 2026. ↗](https://microsoft.ai/news/introducing-mai-cyber-1-flash-inside-mdash/)
[3] Microsoft Security Blog, “Defense at AI speed: Microsoft’s new multi-model agen… ↗
[[4] Wang et al., “CyberGym: Evaluating AI Agents’ Real-World Cybersecurity Capabili… ↗](https://openreview.net/pdf?id=2YvbLQEdYt)
[[5] CyberGym, public benchmark description and leaderboard, accessed July 28, 2026. ↗](https://www.cybergym.io/cybergym/)
[6] Microsoft Security Blog, “How Nationwide stays ahead of attackers with Project … ↗+2 more
The stories that matter, in one email. Free — unsubscribe anytime.