Microsoft has introduced MAI-Cyber-1-Flash, its first dedicated cybersecurity model, and integrated it into MDASH, a vulnerability-hunting system that the company says outperformed Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol while reducing costs by 50% compared with Microsoft’s current best MDASH configuration.

According to Microsoft, the combined MDASH setup scored 95.95% on CyberGym, a benchmark that asks AI agents to reproduce 1,507 known vulnerabilities across 188 open-source projects. CyberGym scores systems by the percentage of vulnerabilities they successfully reproduce in a controlled environment.

Microsoft said the result put MDASH ahead of GPT-5.5 Cyber at 85.6%, Mythos 5 at 83.8%, GPT-5.6 Sol at 83.6%, and Gemini 3.5 Flash Cyber at 83.2%. The score was self-reported by Microsoft and had not appeared on CyberGym’s public leaderboard at the time of publication, although the benchmark uses a public test set and a defined success metric.

Benchmarks such as CyberGym matter because they test whether an agent can move beyond flagging suspicious code and actually reproduce a known flaw. That is still narrower than proving how a system will perform across private codebases, but it gives researchers and security teams a common way to compare vulnerability-hunting systems.

MAI-Cyber-1-Flash is not operating by itself inside the system. Microsoft said the model handles up to 90% of tasks, while MDASH routes the most difficult 10% to GPT-5.4. The routing is significant because tokens—the chunks of text an AI model processes—carry a cost each time a model reads code or generates an answer.

Microsoft described the release as the first time one of its efficiency-focused models has been able to beat a dense, state-of-the-art model designed for broader general capabilities. “When combined with MDASH, (MAI-Cyber-1-Flash) delivers world-class performance at 50 percent of the cost of leading models,” Microsoft CEO Satya Nadella wrote.

Today, we are announcing a series of updates that give customers frontier-grade security at half the cost. MAI-Cyber-1-Flash is our first cybersecurity model, built ground up to find the most challenging vulnerabilities in complex code bases. When combined with MDASH, it… pic.twitter.com/npcIihN1H7 — Satya Nadella (@satyanadella) July 27, 2026

A model, in this case MAI-Cyber-1-Flash, is the AI system that reasons over code. A harness, in this case MDASH, is the surrounding machinery: the agents, tools, checks, and workflow that decide where to look, challenge suspected findings, remove duplicates, and prove that a bug can be triggered.

MDASH uses more than 100 specialized agents assigned to audit code, debate whether a finding is genuine, and build a proof of concept—a working demonstration that the flaw exists. That proof-of-concept step is central for software teams because unverified bug reports can create extra triage work, while reproducible findings are easier to prioritize and fix.

Microsoft said occasional scans and delayed patches are becoming obsolete as AI makes bug discovery cheaper. The company also argued that its decades of security data give it an advantage, saying, “No one can manufacture this history.”

Since the release of Claude Mythos, cybersecurity experts have been trying to match or surpass its capabilities. Researchers reproduced Mythos-style vulnerability hunting with public models for under $30 per scan. Dawid Moczadło, one of the researchers involved, said “the moat is moving from model access to validation.”

That means the scarce part of the process is increasingly the system that can prove findings without overwhelming developers with false alarms.

Decrypt previously reported that GPT-5.5 Cyber had recently taken the public CyberGym lead with a score of 85.6%, narrowly beating Mythos. Microsoft’s newly reported score is about 10 points higher than GPT-5.5 Cyber and 7.5 points above MDASH’s previous result, but it evaluates the full system rather than MAI-Cyber-1-Flash alone.

Microsoft is placing MDASH into private preview through Microsoft Security Exposure Management in the Defender portal. Customers can scan Git repositories, review findings ranked from unlikely to proven, and use the Defender CLI to generate proposed code fixes for developer review.

The preview currently limits repositories to roughly 256MB and allows one concurrent scan per tenant. Project Perception is expected to extend the same multi-agent approach beyond code scanning into broader threat monitoring and remediation workflows.