Microsoft previews Project Perception for continuous agentic security testing

Microsoft described a closed-loop system that combines signals and organisational context with specialised red, blue and green agents. It routes tasks among models according to quality, reliability, latency and cost, while retaining human controls. MAI-Cyber-1-Flash will be used within the MDASH hacking agent.

Why it matters: This moves multi-agent orchestration from general productivity into continuous security validation, where auditability, permissions and failure recovery are essential. Microsoft reports 96% on CyberGym and nearly 50% lower cost than the current MDASH system; those benchmark and cost claims are not independently reproduced.

Impact: Security leaders gain a testable model for agent segregation and model routing, but should require sandboxing, approval gates and their own baselines before production use. Microsoft announcement

AI infrastructure · United States