Tech • IA • Crypto
Microsoft’s New AI Beats Mythos and Shocks OpenAI
9/10TL;DR
Microsoft’s multi-agent AI security system Mdash outperformed top single models on a major benchmark and has already uncovered critical Windows vulnerabilities.
KEY POINTS
Benchmark-Leading Performance
Microsoft’s Mdash scored 88.45% on the CyberGym benchmark, surpassing Anthropic’s Mythos at 83.1% and OpenAI’s GPT-5.5 at 81.8%. The result is notable because competitors relied on their most advanced proprietary models, while Microsoft used generally available models integrated into a broader system. The benchmark includes 157 real-world vulnerability reproduction tasks drawn from 188 open-source fuzzing projects.
Multi-Agent Architecture
Mdash operates as a coordinated pipeline of over 100 specialized AI agents, rather than a single model. These agents perform distinct roles such as auditing, debating, validating, deduplicating, and proving vulnerabilities. The system processes code through five stages: prepare, scan, validate, deduplicate, and prove, enabling layered analysis and verification.
Model-Agnostic Design
The system is designed to be model-agnostic, allowing engineers to swap in newer AI models without rebuilding the pipeline. Different stages use different model types, from high-reasoning systems to smaller, efficient verification models. Disagreements between agents are treated as signals, improving confidence in identified vulnerabilities.
Real Vulnerabilities Discovered
When deployed on Windows code, Mdash identified 16 vulnerabilities, including four critical flaws slated for fixes in a recent Patch Tuesday update. Some of these vulnerabilities could enable remote code execution without authentication, posing significant security risks.
Complex Bug Detection
One discovered flaw, CVE-2026-333827, involved a use-after-free bug in the Windows TCP/IP stack, where memory is accessed after being released. Another, CVE-2026-333824, was a double-free vulnerability spread across six files, enabling potential system-level compromise through just two crafted network packets. These issues required cross-file reasoning and pattern comparison beyond single-function analysis.
High Recall on Known Exploits
Internal testing showed strong performance on historical vulnerabilities. Mdash achieved 96% recall across 28 cases in one Windows component and 100% recall across seven cases in another. On a private test driver with 21 injected vulnerabilities, the system identified all of them with zero false positives.
Limits and Failure Cases
Analysis of failed benchmark tasks revealed that 82% of misidentifications stemmed from vague vulnerability descriptions lacking clear code references. In other cases, the system generated correct exploit logic but failed due to mismatched input formats, highlighting the importance of precise task definitions.
Shift in AI Competition
The results suggest a shift from focusing solely on model strength to emphasizing system engineering and orchestration. Microsoft’s approach demonstrates that combining multiple models effectively can rival or exceed the performance of cutting-edge standalone systems.
Implications for Cybersecurity
The same techniques enabling defenders to find vulnerabilities faster could also be used by attackers. Because Mdash relies on publicly available models and modular design, similar systems could be replicated, accelerating both offensive and defensive cyber capabilities.
Early Deployment փուլ
Mdash is currently in limited private preview, with no announced pricing or general release timeline. Development involved teams from Microsoft’s Autonomous Code Security, Offensive Research and Security Engineering, and Windows Attack Research groups, including members of the DARPA AI Cyber Challenge-winning Team Atlanta.
CONCLUSION
Microsoft’s Mdash highlights a growing shift toward multi-agent AI systems in cybersecurity, showing that orchestration and engineering may rival raw model power while accelerating both defense and attack capabilities.