MAI-Cyber-1-Flash

Microsoft's first in-house cybersecurity model is a 137B sparse MoE fine-tune that drives its MDASH vulnerability harness to a self-reported 95.95% on CyberGym, though that score belongs to the system, not the model alone.

MAI-Cyber-1-Flash

Overview

MAI-Cyber-1-Flash is Microsoft's first purpose-built cybersecurity model, launched July 27, 2026 with Project Perception, an agentic security platform built around it. The model is a security-specialized fine-tune of MAI-Code-1-Flash, which was itself developed from a mid-training checkpoint of MAI-Thinking-1 - so this is the third model in Microsoft's in-house lineage to ship this year, not a clean-sheet build. It runs a sparse Mixture-of-Experts transformer at 137 billion total parameters with 5 billion active per token, a 256K token context window, and text-only input/output.

The model exists to power MDASH, Microsoft's multi-agent vulnerability identification and remediation harness. MAI-Cyber-1-Flash handles roughly 90% of MDASH's workload directly - inventory scanning, log triage, CVE matching, candidate-flaw generation - and escalates the hardest 10% of cases to OpenAI's GPT-5.4, a model Microsoft still licenses under its existing OpenAI agreement. That's notable on its own: Microsoft's flagship in-house security push still leans on a competitor's model for its top-tier reasoning.

TL;DR

  • Best at cheaply handling high-volume triage inside MDASH, escalating only the hardest ~10% of cases to GPT-5.4
  • 137B total / 5B active sparse MoE, 256K context, available only inside MDASH via gated Azure AI Foundry preview - no public API or pricing
  • Headline claim is a system-level 95.95% on CyberGym Level 1 (MAI-Cyber-1-Flash + GPT-5.4 inside MDASH), not a standalone-model score, and it hadn't appeared on CyberGym's public leaderboard as of July 28, 2026

Microsoft is positioning MAI-Cyber-1-Flash and Project Perception as a direct answer to Claude Mythos 5, OpenAI's GPT-5.5-Cyber, and Google's cyber-tuned Gemini offerings, all of which shipped restricted cybersecurity models earlier in 2026. It also follows through on the restructuring reported in mid-July, when security chief Hayete Gallot replaced eight executives and consolidated engineering teams under one reporting line - a reorg that, in hindsight, was clearly building toward this specific launch.

Key Specifications

SpecificationDetails
ProviderMicrosoft
Model FamilyMAI
Base ModelFine-tuned from MAI-Code-1-Flash, which derives from a MAI-Thinking-1 mid-training checkpoint
Parameters137B total, 5B active (sparse MoE)
Context Window256K tokens
ModalityText-only input and output
PricingNot disclosed as a standalone rate - consumption-based via Security Compute Units (SCUs) at the MDASH/Project Perception level, numeric SCU rates not yet published
Release DateJuly 27, 2026
LicenseProprietary - gated Azure AI Foundry preview, MDASH-only
AvailabilityVerified defenders via MDASH; no public API, no self-serve access
Training DataMicrosoft's internal exploit and remediation records, plus commercially licensed data

The pricing line is the one to watch. Microsoft says the new MDASH configuration costs about 50% less than its prior best setup (GPT-5.4 plus GPT-5.4 mini plus GPT-5.3 Codex), but it hasn't published the token volumes, call counts, or task mix behind that comparison - so there's no way to independently verify the savings claim from the outside.

Benchmark Performance

This is where the model card needs the most unpacking, because Microsoft's headline number and the model's actual standalone capability are two very different things.

The Headline Number Is a System Score

Microsoft's marketing centers on a 95.95% result on CyberGym Level 1 - a benchmark where an agent is given a described, known vulnerability plus the pre-patch codebase and has to produce a working proof-of-concept exploit. It's a reproduction and triage task, not blind vulnerability discovery. That distinction matters for reading the number correctly.

The 95.95% score belongs to MDASH running MAI-Cyber-1-Flash with GPT-5.4 - not to MAI-Cyber-1-Flash by itself.

SystemCyberGym Level 1Source
MDASH (MAI-Cyber-1-Flash + GPT-5.4)95.95%Microsoft, self-reported July 27, 2026
Wiz Atlas90.9%CyberGym public leaderboard, as of July 28, 2026
MDASH, prior configuration (GPT-5.4 + GPT-5.4 mini + GPT-5.3 Codex)88.45%Microsoft, May 2026 submission
GPT-5.5-Cyber85.6%OpenAI
Claude Mythos 583.8%Anthropic
GPT-5.6 Sol83.6%OpenAI

As of July 28, 2026, Microsoft's 95.95% figure hadn't been posted to CyberGym's own public leaderboard, which still showed Wiz's Atlas agent leading at 90.9% and Microsoft's own May submission sitting at 88.4%. The 95.95% claim comes exclusively from Microsoft's launch materials. That doesn't make it false, but it does mean it hasn't cleared the same independent-verification bar the leaderboard numbers have. Until Microsoft submits the new configuration for public listing, treat the comparison to Mythos 5, GPT-5.5-Cyber, and GPT-5.6 Sol as Microsoft's own framing, not a neutral third-party ranking.

Standalone Model Numbers Are Much Weaker

Reporting from The Hacker News surfaced the model's scores on harder, non-CyberGym benchmarks, run under a lightweight terminal harness without MDASH's orchestration or GPT-5.4 escalation:

BenchmarkMAI-Cyber-1-Flash (standalone)
CVEBench0.314
CyberSecEval4 (threat intelligence)0.553
Malware analysis0.33
CRSBench (POV=1200)0.651
ExploitGym (kernel / userspace / browser)0.0 across all three categories

A zero across every ExploitGym category is a real data point, not noise - it means the standalone model currently can't produce a working exploit chain on that benchmark's harder tasks at all. Read together with the CyberGym system score, the picture is consistent: MAI-Cyber-1-Flash is a capable, cheap triage-and-reproduction model when it has GPT-5.4 and MDASH's orchestration to lean on, and a much weaker one operating alone against tasks that require deeper exploit development. That's exactly the shape of model Microsoft's own architecture implies - it was built to handle volume, not to be the reasoning engine at the top of the stack.

Key Capabilities

High-Volume Triage Inside MDASH

MDASH runs security work through five stages - Prepare, Scan, Validate, Dedupe, and Prove - coordinating more than 100 specialized agents. MAI-Cyber-1-Flash aims to be the workhorse across most of that pipeline: inventory checks, log parsing, CVE matching, deduplication of candidate findings, and a first pass at vulnerability validation. Microsoft says it draws on "decades of security history and an unmatched record of real exploits and remediations" from its own internal vulnerability database - a training-data advantage that's plausible given Microsoft's scale as a target, though the company hasn't published details on dataset composition or size.

Escalation to GPT-5.4 for the Hard 10%

The architecture's most honest signal is the escalation path itself. Rather than pushing MAI-Cyber-1-Flash to handle everything, MDASH routes the roughly 10% of cases that require deeper reasoning - reconstructing an exploit chain, interpreting an unusual authentication flow - to GPT-5.4. A Microsoft executive was blunt about the reasoning: "GPT-5.4 is incredibly good relative to its cost," and "the whole game here is to reduce the costs." That's a cost-engineering statement, not a capability claim for the new model.

Close-up of a padlock securing a metal cabinet, representing the layered access controls behind MDASH MAI-Cyber-1-Flash never ships as a standalone API - access runs completely through MDASH's gated, sandboxed environment.

Enterprise Controls and Sandboxing

MDASH deployments include role-based access control, tenant isolation, encryption, auditability, and sandboxed execution with no internet connectivity for the agents doing the actual exploit-generation work. Microsoft says the model went through internal AI Red Team review, adversarial testing, and third-party assessment before launch - standard practice for a model whose core function (writing working exploit code) is dual-use by design.

A close-up of colorful syntax-highlighted source code on a screen Reproducing a CyberGym Level 1 case means turning a vulnerability description and pre-patch source into a working proof-of-concept - a narrower task than open-ended vulnerability discovery. Source: pexels.com

Pricing and Availability

MAI-Cyber-1-Flash isn't sold as a standalone model. It's available exclusively through MDASH, in a gated Azure AI Foundry preview restricted to vetted enterprise customers - Microsoft cites unspecified "Fortune 500" deployments already running it. There's no self-serve signup; interested customers submit a registration form and go through approval, a deliberate friction point given that vulnerability discovery and exploit validation tooling is exactly as useful to attackers as to defenders.

Pricing sits one layer up, at the Project Perception platform level, using consumption-based Security Compute Units (SCUs). Different agents draw SCUs at different rates depending on task intensity, but Microsoft hasn't published numeric SCU rates as of this writing. The company's cost claim - roughly 50% cheaper than the prior best MDASH configuration - is the only concrete pricing signal available, and it's a relative claim, not an absolute price.

  • April 2, 2026 - Microsoft ships its first three MAI models (transcription, TTS, image), establishing the in-house model program.

  • June 2, 2026 - MAI-Thinking-1 and MAI-Code-1-Flash launch at Build 2026, the reasoning and coding foundations MAI-Cyber-1-Flash builds on.

  • July 17, 2026 - Reporting surfaces on Project Perception and the security-org restructuring under Hayete Gallot, ahead of any official confirmation.

  • July 27, 2026 - MAI-Cyber-1-Flash and Project Perception launch officially, with the 95.95% CyberGym claim as the headline number.

  • August 3, 2026 - Project Perception enters public preview, starting inside Microsoft Defender before expanding across the security portfolio.

Project Perception itself - the surrounding agentic platform with red team (attack simulation), blue team (detection and triage), and green team (remediation) agents, coordinated through a six-layer "Cyber Stack" - enters public preview on August 3, 2026. That's a separate product from the model covered here, but it's the only channel most customers will actually touch MAI-Cyber-1-Flash through.

Strengths and Weaknesses

Strengths

  • Handles ~90% of MDASH's task volume at a fraction of the cost of routing everything to a frontier model
  • 256K context window and MoE efficiency (5B active of 137B total) make high-volume scanning economically viable
  • Built on Microsoft's internal exploit/remediation history, a training-data source competitors can't copy
  • Enterprise-grade sandboxing (no internet access for exploit-generation agents, RBAC, tenant isolation) addresses the dual-use risk directly
  • Part of a coherent three-model lineage (Thinking-1 to Code-1-Flash to Cyber-1-Flash) rather than a one-off fine-tune

Weaknesses

  • The 95.95% headline score is a system-level number (with GPT-5.4 escalation), not a standalone model result - don't compare it directly to other vendors' standalone cyber-model scores
  • Not listed on CyberGym's public leaderboard as of July 28, 2026 - the claim is currently self-reported only
  • Standalone scores on harder benchmarks are weak: 0.314 on CVEBench, 0.33 on malware analysis, and a flat zero across every ExploitGym category
  • No standalone API, no public pricing, no self-serve access - evaluation requires a vetted MDASH relationship with Microsoft
  • The cost-savings claim (50% cheaper) lacks published token volumes or task-mix data, so it can't be independently checked
  • Still architecturally dependent on OpenAI's GPT-5.4 for the hardest 10% of cases, undercutting the "independent security stack" narrative

FAQ

Is MAI-Cyber-1-Flash available to the public?

No. It only runs inside MDASH, Microsoft's vulnerability harness, through a gated Azure AI Foundry preview restricted to vetted enterprise customers. There's no public API, waitlist, or self-serve pricing.

Does MAI-Cyber-1-Flash really score 95.95% on CyberGym?

That score belongs to MDASH running MAI-Cyber-1-Flash alongside OpenAI's GPT-5.4, not to the standalone model. As of July 28, 2026, it also hadn't appeared on CyberGym's public leaderboard, which still listed Wiz's Atlas agent at 90.9%.

What does CyberGym Level 1 actually measure?

Given a vulnerability description and the pre-patch source code, an agent must create a working proof-of-concept that reproduces a known, real-world vulnerability. It's a reproduction and triage task, not open-ended vulnerability discovery.

How does MAI-Cyber-1-Flash compare to Claude Mythos 5 or GPT-5.5-Cyber?

On Microsoft's self-reported system-level CyberGym number, it leads both. On standalone benchmarks like CVEBench and ExploitGym, Microsoft's own disclosed numbers are considerably weaker, and no independent standalone comparison exists yet.

Why does MDASH still rely on GPT-5.4?

Microsoft says GPT-5.4 offers strong reasoning relative to its cost for the hardest ~10% of tasks. MAI-Cyber-1-Flash is built to absorb high-volume triage work cheaply, not to replace frontier-tier reasoning models outright.

When does Project Perception launch?

Project Perception, the agentic platform built around MAI-Cyber-1-Flash and MDASH, enters public preview on August 3, 2026, starting inside Microsoft Defender.


Sources:

✓ Last verified July 28, 2026

James Kowalski
About the author AI Benchmarks & Tools Analyst

James is a software engineer turned tech writer who spent six years building backend systems at a fintech startup in Chicago before pivoting to full-time analysis of AI tools and infrastructure.