Microsoft has introduced MAI-Cyber-1-Flash, its first cybersecurity-specific artificial intelligence model, integrated into the MDASH platform. This development aims to bolster vulnerability identification and remediation processes.
In recent evaluations, MDASH, utilizing MAI-Cyber-1-Flash alongside GPT-5.4, achieved a 95.95% score on CyberGym Level 1. This configuration reportedly operates at half the cost of the previous MDASH setup, which combined GPT-5.4, GPT-5.4 mini, and GPT-5.3 Codex. Access to this enhanced system is currently restricted to approved MDASH customers through an Azure AI Foundry private preview.
MAI-Cyber-1-Flash is engineered to manage up to 90% of MDASH tasks, reserving GPT-5.4 for the most complex 10%. Notably, MAI-Cyber-1-Flash is exclusive to MDASH and is not available as a standalone model or through a general-purpose API.
It’s important to note that the 95.95% score pertains to the combined performance of MDASH running both MAI-Cyber-1-Flash and GPT-5.4, rather than MAI-Cyber-1-Flash alone. CyberGym Level 1 assesses an agent’s ability to reproduce known vulnerabilities by providing a vulnerability description and the corresponding unpatched source code, then evaluating whether the agent can produce a working proof of concept. This test does not measure the agent’s capability in blind vulnerability discovery or the accuracy of generated patches.
As of July 28, 2026, CyberGym’s public leaderboard did not list Microsoft’s 95.95% result. The top position was held by Wiz’s Atlas agent at 90.9%, with Microsoft’s previous MDASH submission from May 12 standing at 88.4%. Microsoft’s public materials do not specify whether the new result was submitted for leaderboard inclusion.
Microsoft’s earlier MDASH result of 96.55% from June accounted for any crash, including those from non-target vulnerabilities. The July materials do not clarify if the 95.95% result follows the same criteria, making direct performance comparisons challenging.
According to Microsoft’s model card, MAI-Cyber-1-Flash is a sparse mixture-of-experts transformer with 137 billion total parameters, five billion active parameters, and a 256,000-token context window. It is a cybersecurity fine-tune of MAI-Code-1-Flash, developed from a MAI-Thinking-1 mid-training checkpoint.
The model card indicates that the evaluated configuration replaced 80% of MDASH’s existing models, elevating the reported CyberGym result from 88.4% to 95.95%. The 80% figure represents the proportion of models replaced, while the 90% figure denotes the maximum share of tasks the smaller model is designed to handle.
Microsoft’s launch announcement defines the 50% cost saving in comparison to its current best MDASH model mix of GPT-5.4, GPT-5.4 mini, and GPT-5.3 Codex. The product page describes the system as delivering “comparable performance at 50% of the cost of leading models.” However, the announcement and model card do not disclose specific details regarding token usage, call volume, latency, task mix, or compute allocation behind this comparison, making independent reproduction or normalization against other systems currently unfeasible.
Taesoo Kim, Microsoft’s vice president of agentic security, emphasized the distinction between the model and the system, stating, “The model is one input, the system around it is the product.” Under a lightweight terminal harness, the model card reports scores of 0.314 on CVEBench, 0.553 on CyberSecEval4 threat intelligence, 0.33 on its malware-analysis test, and 0.651 on CRSBench at POV=1200.
Interestingly, the model scored zero across the kernel, userspace, and browser categories of ExploitGym, which challenges agents to turn supplied vulnerabilities and crashing inputs into functional exploits.
Microsoft’s integration of MAI-Cyber-1-Flash into MDASH signifies a strategic move to enhance cybersecurity capabilities through AI. By achieving higher accuracy at reduced costs, this development could set a new standard in vulnerability management. However, the lack of transparency regarding certain evaluation metrics and the absence of the new score on public leaderboards suggest that further validation and peer review are necessary to fully assess the model’s impact and reliability.
