May 13, 2026 — The AI security landscape has a new entrant: Microsoft’s Autonomous Code Security (ACS) team has officially launched MDASH (Multi-Model Agentic Scanning Harness). The framework scored a high 88.45% on the authoritative CyberGym public benchmark, decisively outperforming Anthropic’s Claude Mythos Preview (83.1%) and OpenAI’s GPT-5.5 (81.8%), and taking the top spot. This result not only marks a major Microsoft breakthrough in AI-driven code security, but also highlights the clear advantages of “multi-model agent collaboration” over a single leading model.

MDASH was developed by Microsoft’s security team and is already deployed in production. It helped Microsoft identify and remediate 16 Windows vulnerabilities during a recent Patch Tuesday, including four high-severity remote code execution (RCE) flaws affecting critical components such as the TCP/IP kernel stack, IKEv2, the Netlogon service, and the DNS API library.

MDASH Implementation Details: 100+ Specialized Agents Working in Concert

Unlike single-model solutions such as Anthropic Mythos and OpenAI GPT-5.5-Cyber, MDASH uses a multi-model agent pipeline architecture with the following core attributes:

More than 100 specialized AI agents: these agents operate across different stages, covering code comprehension, vulnerability hypothesis generation, static analysis, dynamic verification, adversarial debate, and exploitability proof-of-concept (PoC) generation.
Hybrid model integration: it invokes both frontier models (e.g., the latest GPT/Claude series) and Microsoft’s own distilled models, using dynamic routing to combine their strengths.
End-to-end agentic workflows: agents do not only find vulnerabilities but also generate working PoC exploit code and employ multi-round “debate” mechanisms to reduce false positives. The system emphasizes reproducibility and aims for zero false positives (internally achieving 21/21 injected-vulnerability zero false positives).
Tight engineering integration: it is deeply integrated with Microsoft’s existing code repositories, fuzzing infrastructure, and security scanning pipelines to deliver “defense at AI speed.”

The CyberGym benchmark, developed by UC Berkeley, contains 1,507 real-world vulnerability tasks drawn from 188 open-source projects (based on Google OSS-Fuzz). Tests require the AI system to produce a PoC that successfully triggers a vulnerability given the vulnerability description and an unpatched codebase. This benchmark closely mirrors real offense–defense scenarios and is among the strictest public standards for evaluating AI vulnerability discovery capabilities.

Microsoft’s security blog highlights that MDASH also shows very high recall on historical MSRC (Microsoft Security Response Center) cases, reaching 96%–100% detection capability on components such as clfs.sys and tcpip.sys.

CyberGym benchmark scores over time, showing the rapid improvement in AI vulnerability discovery capabilities. Microsoft’s multi-model MDASH system (top right) tops the leaderboard at 88.4%. (CyberGym / UC Berkeley)

Implications for Code Security and Industry

1. Multi-agent architectures demonstrate clear advantages

MDASH’s success shows that for complex security tasks, a single oversized model is not necessarily optimal. By dividing labor, employing specialized agents, and integrating multiple models,