Microsoft has announced Project Perception – an AI-powered platform that brings together dozens of specialised agents to automatically detect and eliminate cyber threats. The New York Times reports on the new tools.
As explained by Microsoft, traditional cyber defence systems can no longer keep up with automated attacks capable of creating exploits and scaling campaigns more quickly.
Project Perception uses three types of specialised agents. The 'red' team looks for potential points of entry before attackers can exploit them. The 'blue' team investigates the issues found and assesses their actual risk, whilst the 'green' team patches vulnerabilities and strengthens defences. Humans retain control over the system's operation.
The first scenario is based on the new MAI-Cyber-1-Flash model, which Microsoft has integrated into the MDASH platform. It utilises over 100 agents trained to detect, verify and fix software vulnerabilities.
MAI-Cyber-1-Flash handles up to 90 per cent of tasks, whilst the system passes the most complex 10 per cent on to the more powerful and expensive GPT-5.4. This division of labour is intended to reduce costs compared to using large models for every check.
According to Microsoft's results, the combination of MAI-Cyber-1-Flash and GPT-5.4 achieved a score of 95.95 per cent in the CyberGym test, which assesses the ability to detect real-world vulnerabilities in large code repositories. This is approximately 12 percentage points higher than the result achieved by Mythos from Anthropic. At the same time, the new configuration costs almost half as much as the version of MDASH, which utilised GPT-5.4, GPT-5.4 mini and GPT-5.3 Codex. Microsoft has not published any independent verification of these results.
Project Perception is not limited to finding errors in code. The system gathers context an organisation's accounts, devices, applications, data, cloud infrastructure and AI services. For training and analysis, Microsoft uses over 100 trillion daily security signals and information from 1.6 million customers.
The agents operate in environments without internet access, and Microsoft promises role-based access control, isolation of different customers' data, encryption and activity logging. The system is designed not only to generate alerts but also to immediately link a detected issue to a specific mitigation action.
Public testing of Project Perception will begin on 3 August 2026. Microsoft has announced a global launch, but has not yet specified the price or the list of supported regions.
As a reminder, in May, the MDASH system helped Microsoft researchers to discover 16 new vulnerabilities in Windows networking components and authentication systems. Four of these allowed remote code execution and were classified as critical.
The full story
This article is one source in a clustered incident — the cluster page carries the summary, timeline and every other outlet covering it.
