Microsoft's MDASH AI system has found 16 previously unknown vulnerabilities in Windows. The system, which scored 88.45% on the CyberGym benchmark, outperformed both Anthropic's Mythos and OpenAI's models in security bug detection.
What the CyberGym benchmark measured
CyberGym is a standardized test for AI security tools. MDASH AI's 88.45% score places it ahead of Anthropic's Mythos and OpenAI's systems in finding vulnerabilities. The benchmark simulates real-world attack scenarios, testing how well AI can spot flaws in code.
Outperforming rival AI systems
Anthropic's Mythos and OpenAI's models are among the most advanced AI systems for security tasks. But MDASH AI beat them both in this round of testing. The 16 vulnerabilities it found are all new — meaning they weren't previously known to Microsoft or the security community.
What happens next
Microsoft now has the list of vulnerabilities. The company hasn't said when it will release patches, but the discovery gives its security team a head start on fixing the flaws before attackers can exploit them. The results also highlight how AI is becoming a bigger part of the cybersecurity toolkit.
