How does Project Perception organize its AI agents?

The platform creates three teams: red agents that simulate attackers, blue agents that investigate and rank alerts, and green agents that write and deploy patches, while still requiring human approval for critical actions.

What performance did the new system achieve on the CyberGym benchmark?

MDASH powered by MAI-Cyber-1-Flash scored about 96% on the public CyberGym benchmark, surpassing Anthropic’s Mythos (≈84%) and earlier MDASH results.

How much cost reduction does Microsoft say enterprises can expect?

Microsoft says the multi-model agentic approach can cut security operating costs by roughly 50% compared with using only large frontier models.

When did Project Perception become available for testing?

Project Perception entered public preview on August 3, 2026 and is priced based on consumption measured in Security Compute Units.

Microsoft Agentic Security Model Cuts Costs, Beats AI Benchmark

Q: What is Microsoft’s new cybersecurity model called?

It is called MAI-Cyber-1-Flash, a compact model tuned on Microsoft’s own exploit and remediation data that runs inside the MDASH scanning system.

TL;DR: Microsoft launched MAI-Cyber-1-Flash, a custom security model that handles most tasks inside its MDASH scanning system, and Project Perception, an agentic defense platform that groups AI agents into red, blue and green teams. The combined system scores about 96% on the CyberGym benchmark, outpacing leading frontier models, and promises roughly 50% cost savings for enterprises.

Key facts

MDASH (Microsoft Security multi-model agentic scanning harness) coordinated more than 100 specialized AI agents across multiple frontiers and distilled models.
MDASH helped researchers discover 16 previously unknown Windows networking and authentication vulnerabilities, including four critical remote-code-execution bugs that were patched in May’s Patch Tuesday release.
MDASH achieved an 88.45% score on the CyberGym benchmark, which evaluates AI reasoning over more than 1,500 real-world vulnerabilities.
MAI-Cyber-1-Flash, a compact code-tuned derivative of the MAI-Thinking-1 line, runs inside MDASH and handles roughly 90% of security workload, with the remaining 10% routed to OpenAI’s GPT-5.4, cutting operating costs by about half.
When powered by MAI-Cyber-1-Flash, MDASH scored 95.95% (reported as 96%) on the public CyberGym benchmark, beating Anthropic’s Mythos (≈84%) and earlier MDASH results.
Project Perception groups agents into three color-coded teams—red (attack simulation), blue (triage), and green (fix generation)—with high-impact actions requiring human sign-off, and entered public preview on August 3, 2026.
Microsoft claims the multi-model approach can deliver roughly 50% cost savings for enterprises compared with relying solely on large frontier models.

Microsoft expands its agentic security strategy

In a two-part rollout announced between May and July 2026, Microsoft introduced a custom-built security model and a new agentic defense platform. The company says the combination halves the cost of running AI-driven security while delivering a near-perfect score on a leading vulnerability-reproduction benchmark.

MDASH: a multi-model agentic scanning harness

MDASH (Microsoft Security multi-model agentic scanning harness) coordinates more than 100 specialized AI agents that autonomously analyze source code, debate whether a finding is exploitable, validate the result, and even generate proof-of-concept exploits [1][4]. Taesoo Kim, Microsoft’s vice president of agentic security, described the workflow as “agents that can discuss, argue, and converge on a reliable exploit without human prompting” [1].

The system helped internal researchers uncover 16 previously unknown Windows networking and authentication vulnerabilities, among them four critical remote-code-execution bugs that were patched in the May Patch Tuesday release [1][4]. The effort was bolstered by contributions from Team Atlanta, the winners of DARPA’s AI Cyber Challenge.

On the CyberGym benchmark, which tests AI reasoning over more than 1,500 real-world vulnerabilities, MDASH originally scored 88.45% [1][4].

MAI-Cyber-1-Flash: Microsoft’s first in-house security model

In July 2026, Microsoft unveiled MAI-Cyber-1-Flash, a compact, code-tuned derivative of the MAI-Thinking-1 family trained on Microsoft’s own exploit and remediation data [3]. The model runs inside MDASH and handles roughly 90% of the security workload, while the remaining 10% is delegated to OpenAI’s GPT-5.4 [4]. This split cuts the overall operating cost of the harness by about 50%.

When MDASH operated with MAI-Cyber-1-Flash, its CyberGym score rose to 95.95% (VentureBeat reported a rounded 96%), outperforming Anthropic’s Mythos (≈84%) and the earlier MDASH configuration [3][4]. The result demonstrates that a purpose-built, smaller model can beat larger frontier models when paired with a coordinated agentic system.

Project Perception: an agentic defense platform

Alongside the model, Microsoft launched Project Perception, an agentic defense platform that groups AI agents into three color-coded teams:

Red agents simulate attacker behavior and probe for weaknesses.
Blue agents investigate alerts, rank their severity, and triage the most urgent signals.
Green agents write and deploy remediation code.

High-impact actions still require human sign-off, keeping ultimate control with security operators [3][4]. The platform is priced on consumption using Security Compute Units and entered public preview on August 3, 2026.

Hayete Gallot, EVP of Microsoft Security, said Project Perception “brings together signals, context, models and specialized agents into a continuously learning system of defense” [3]. Microsoft AI CEO Mustafa Suleyman added that the architecture delivers “faster, better, cheaper models” and represents “the tip of the iceberg” for future security AI [4].

Why the multi-model, agentic approach matters

Microsoft’s broader “agentic era” vision envisions autonomous AI agents woven into every layer of the security stack—from silicon to cloud services—to provide ambient, always-on protection [2]. In this view, agentic AI refers to systems that can plan, reason, and act with minimal human intervention [5][6]. By combining a specialized model (MAI-Cyber-1-Flash) with a large-scale, multi-agent orchestration (MDASH and Project Perception), Microsoft claims enterprises can achieve roughly 50% cost savings compared with relying solely on large frontier models [4].

Outlook

If the early results hold up, Microsoft’s agentic security stack could become a blueprint for other vendors seeking cost-effective, high-performance AI defenses. The public preview of Project Perception will allow organizations to test the platform’s real-world impact, while the open-source community watches to see whether the multi-model approach can consistently outpace ever-larger foundation models.

Sources

Microsoft Intros New Agentic AI Security Multi-Model Defense System — Campus Technology (campustechnology.com) — 2026-05-18
Microsoft launches AI cybersecurity model, agentic defense platform to cut enterprise security costs (venturebeat.com) — 2026-07-27
Microsoft’s introduces its first agent-powered cybersecurity model - SiliconANGLE (siliconangle.com) — 2026-07-27
Microsoft Ignite: Ambient and autonomous security for the agentic era | Microsoft Security Blog (www.microsoft.com) — 2025-11-18
What Is Agentic AI in Cybersecurity? | Microsoft Security (www.microsoft.com) — 2026-04-09
What Is Agentic AI Security? | Microsoft Security (www.microsoft.com) — 2026-07-01