Microsoft Launches Project Perception: 100+ AI Agents Team Up to Hunt Vulnerabilities, Public Preview August 3
Microsoft just turned cybersecurity into an AI-vs-AI war game — its new Project Perception platform deploys over 100 specialized AI agents organized into red, blue, and green teams to automatically discover, validate, and patch software vulnerabilities.
Three Key Takeaways
A three-team agentic architecture. Unveiled on July 27 in San Francisco, Project Perception enters public preview on August 3. The platform's core design features three classes of collaborating agents: "red team" agents proactively search for attack paths, "blue team" agents investigate and triage risks, and "green team" agents handle remediation and defense hardening. These agents share security intelligence, mirroring how human security teams divide labor.
A purpose-built model: MAI-Cyber-1-Flash. Microsoft simultaneously released MAI-Cyber-1-Flash, a model trained specifically for cybersecurity scenarios. According to Microsoft's published benchmarks, when paired with OpenAI's GPT-5.4, the model outperforms Anthropic's Mythos 5, Google's Gemini 3.5 Flash Cyber, and OpenAI's GPT-5.5 Cyber on the CyberGym benchmark. Microsoft emphasized that the solution costs less than competing offerings.
MDASH: a 100+ agent vulnerability hunter. As part of Project Perception, the MDASH system coordinates over 100 specialized AI agents and multiple AI models to analyze codebases, validate threats, and determine whether a vulnerability is actually exploitable in real-world conditions. Microsoft positions it as a proactive defense tool that finds exploitable vulnerabilities before attackers do.
WangDou's Take
Microsoft's logic for selling security products has always been straightforward: Windows is the most attacked operating system on the planet, so Microsoft selling security software is like a fire chief selling extinguishers — you can hardly say no. But what makes Project Perception interesting is the "red team agent" design: Microsoft built a batch of AI agents specifically to attack its own customers' systems, then tells those customers, "Look, I found your vulnerabilities — want to buy my blue and green teams to fix them?" This is penetration testing turned into a product. Beating GPT-5.5 Cyber on CyberGym sounds impressive, but the real test in security has never been benchmarks — it is your response time when a zero-day drops. And that is a number no vendor dares to publish.
Source: Microsoft Official Blog, VentureBeat, Axios