Overview
- Microsoft announced Monday that it has launched MAI‑Cyber‑1‑Flash inside its MDASH multi‑agent harness and introduced Project Perception, an orchestration platform that routes tasks among specialist agents and models.
- The company reports the MDASH configuration using MAI‑Cyber‑1‑Flash (with GPT‑5.4 for the hardest cases) scored about 95.9–96% on the CyberGym vulnerability benchmark but the result is self‑reported and was not listed on CyberGym’s public leaderboard at the time of reporting.
- Microsoft says MAI‑Cyber‑1‑Flash handles roughly 90% of MDASH tasks while MDASH routes the toughest 10% to larger models, a design the company says yields about a 50% reduction in token and compute cost versus its prior MDASH mix.
- MDASH has already been used to find 16 previously unknown Windows vulnerabilities that Microsoft patched in May 2026, and Project Perception will enter public preview on August 3 with MAI‑Cyber‑1‑Flash available to approved customers in a private preview through Azure AI Foundry.
- Microsoft emphasizes layered safety controls including AI red‑teaming, a third‑party assessment, tenant isolation and role‑based access but independent verification of the benchmark claims and wider governance questions about agentic systems remain unresolved.