Microsoft opens Project Perception to the public, with its own security model MAI-Cyber-1-Flash

Project Perception, Microsoft's agentic security system, entered public preview on 3 August 2026, following the announcement posted by Hayete Gallot, EVP of Microsoft Security, on 27 July. The company's own description of the design is unusually concrete: three classes of agents work in a loop, with red agents mapping attack paths, blue agents judging which of those paths represent real risk, and green agents remediating and hardening. The first scenario shipped is software vulnerability management, and the system runs inside Microsoft Defender at launch, with the rest of the Microsoft Security portfolio to follow. The part that concerns this catalogue is the model underneath. Microsoft states that "MDASH with MAI-Cyber-1-Flash delivers 96% on CyberGym", twelve points above Mythos, at "almost 50% of cost savings vs. the current MDASH configuration in market today". That is a vendor-reported benchmark on a vendor-chosen configuration, and it should be read as such — but the direction is clear enough: Microsoft is putting a security-specific model of its own between the customer and the frontier models it has been buying. Trade coverage adds that MAI-Cyber-1-Flash is Microsoft's first model built specifically for cybersecurity and that it hands roughly a tenth of the hardest cases to GPT-5.4; neither claim appears in Microsoft's own post, so we mark them as press reporting. Billing is consumption-based and pay-as-you-go, denominated in Security Compute Units. Microsoft has not published the rate per unit, which makes the claimed cost saving impossible to check independently. MAI-Cyber-1-Flash does not yet have a profile in wujec.ai. It is a named, first-party model with a published benchmark claim, so it belongs in the catalogue and is going into our queue.