Microsoft opens Project Perception for public preview

Microsoft's Project Perception, an agentic cybersecurity platform announced on 27 July, opened for public preview. The system coordinates specialised teams of AI agents, described as red for attack simulation, blue for threat triage and green for remediation, to run vulnerability-management workflows end to end across enterprise environments.

The preview also introduces MAI-Cyber-1-Flash, Microsoft's first in-house model built specifically for cybersecurity work. Paired with GPT-5.4 inside Microsoft's MDASH agent harness, the combination scored roughly 96 percent on CyberGym, a benchmark for AI hacking and defence capability, which Microsoft said put it about 12 points ahead of Anthropic's Mythos model. Microsoft is offering the configuration at half the cost of its previous best cybersecurity offering.

Handing vulnerability triage to autonomous agents raises the same question agentic coding tools did: how much decision-making enterprises will delegate before a person reviews the result.