MAI-Cyber-1-Flash and the MDASH configuration
Microsoft introduced MAI-Cyber-1-Flash as its first cybersecurity-specific model inside MDASH, the company’s multi-model vulnerability identification and remediation harness. The model is described in the model card as a sparse mixture-of-experts transformer with 137 billion total parameters, five billion active parameters, and a 256,000-token context window. Microsoft says MAI-Cyber-1-Flash is a cybersecurity fine-tune of MAI-Code-1-Flash, developed from a MAI-Thinking-1 mid-training checkpoint.
The published design points to a routing strategy: Microsoft says MAI-Cyber-1-Flash is intended to handle up to 90% of MDASH tasks, with GPT-5.4 reserved for the hardest 10%. The company frames the reported 95.95% figure as belonging to MDASH running MAI-Cyber-1-Flash alongside GPT-5.4, not to the new model by itself. Access to this configuration is limited to approved MDASH customers through an Azure AI Foundry private preview; MAI-Cyber-1-Flash is not available as a standalone public model or general-purpose API.
What the CyberGym result measures — and what it does not
Microsoft reported the 95.95% score on CyberGym, and the company specifically notes that the headline result uses CyberGym Level 1 testing. CyberGym Level 1 is a known-vulnerability reproduction test: it supplies an agent a vulnerability description and the corresponding unpatched source code, then checks whether the agent can produce a working proof of concept. CyberGym Level 1 does not measure blind vulnerability discovery, nor whether a generated patch is correct.
When checked on July 28, 2026, CyberGym’s public leaderboard did not list Microsoft’s 95.95% submission. The leaderboard still displayed Microsoft’s May 12 MDASH submission at 88.4%. Microsoft’s public materials do not say whether the 95.95% result was submitted for public listing on CyberGym. The company’s earlier June announcement of a 96.55% MDASH figure counted any crash, including non-target vulnerabilities; the July materials do not say whether the 95.95% used the same crash-counting criterion, so the two published percentages cannot be read as a simple before-and-after performance trend.
Benchmarks, limits, and the system-level framing
Microsoft’s model card lists additional benchmark outputs under a lightweight terminal harness: 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. The card also reports zero scores across the kernel, userspace, and browser categories of ExploitGym, a different test that asks agents to convert supplied vulnerabilities and crashing inputs into working code-execution exploits. These results derive from different tasks and scoring scales and are not presented as standalone CyberGym scores for MAI-Cyber-1-Flash.
Microsoft framed the achievement as a system-level result. “The model is one input, the system around it is the product,” Taesoo Kim, Microsoft’s vice president of agentic security, said in June when describing MDASH. Consistent with that view, the model card says the evaluated configuration replaced 80% of MDASH’s existing models and raised the reported CyberGym result from 88.4% to 95.95% — a figure that reflects the system’s overall performance, not the new model operating alone.
Cost, reproducibility, and deployment constraints
Microsoft’s launch announcement defines the 50% saving against its current best MDASH model mix of GPT-5.4, GPT-5.4 mini, and GPT-5.3 Codex. The product page similarly describes the system as delivering “comparable performance at 50% of the cost of leading models.” The announcement and model card do not disclose the token consumption, call volume, latency, task mix, or compute allocation behind that comparison, meaning the claimed cost saving cannot yet be independently reproduced or normalized against other systems.
Microsoft says all benchmark testing took place in a network-isolated environment with no access to production systems, the public internet, or external services. The model card also warns that generated text and code may be inaccurate or incomplete and should be reviewed before consequential use.
Project Perception, the MDASH scenario, and the near-term rollout
Software vulnerability management using MAI-Cyber-1-Flash inside MDASH is the first scenario Microsoft has publicly announced for Project Perception, the company’s broader system for coordinating defensive security agents. Project Perception is scheduled to enter public preview on August 3, and Microsoft plans to extend the model beyond software vulnerability work to additional security workflows. For now, Microsoft is keeping the model and the MDASH configuration behind a private-preview gate through Azure AI Foundry for approved customers.
Microsoft’s materials present a clear system-level argument: pair a smaller, specialized model with a larger generalist for the hardest tasks and claim improved accuracy and cost-efficiency at the MDASH level. The record published so far leaves the 95.95% figure anchored to that system claim, to controlled testing in isolation, and to a private-preview deployment; whether the result appears on CyberGym’s public leaderboard or is reproduced under public conditions remains to be seen as Project Perception moves toward public preview.
