Microsoft Perception, a system of cooperating security agents, is now generally available, Satya Nadella announced at Ignite on 17 November 2026. Microsoft company page also put Codename MDASH, an agentic code security system, into preview.
Quick Take
- Perception: generally available; teams of red, blue and green agents run playbooks with humans in control.
- Codename MDASH: preview for eligible organizations; more than 100 AI agents in a four-stage pipeline.
- Model: MAI-Cyber-2-Flash, in preview only inside MDASH, will be integrated with Perception.
- Agent risk: Dynamic Agent Risk in Agent 365 is in preview and uses partner signals from Check Point, Cyera, Darktrace and Zenity.
In plain English
Microsoft is using AI agents to defend against attacks by AI. One group of agents plays the attacker to find weak points, another watches for real attacks, and a third repairs problems. The user chooses the playbooks and stays in charge. MDASH applies the same idea to source code, looking for security flaws and checking them before they reach developers.
What happened
Nadella said: “red team agents simulate attacks, blue team agents detect attacks and green team agents fix and remediate.” He added that with Playbooks “you can compose teams of agents to tackle complex jobs and trigger them to take action autonomously,” and “I am very excited to share that Perception is now generally available.” Perception links signals across identity, client, cloud, apps and agents.
Background
Microsoft says AI advances are changing how vulnerabilities are found and exploited. For a recent example of attacks on coding agents, see our report on device-code phishing aimed at AI coding agents.
How it works
MDASH maps the attack surface, scans code, validates findings through agent debate, removes duplicates and, where possible, reproduces a finding with a triggering input. New in preview are repository sharding to use fewer tokens per scan, progress visibility and wider MAI Cyber model support.
Why it matters
For security teams, fewer false positives and validated findings mean less triage.
For attackers and defenders, machine-speed response is now sold as a product feature.
For rivals, security vendors face a platform with models, agents and identity data in one place.
What’s still open
- Independent results on detection and false positive rates.
- Pricing and licensing for Perception.
- When MDASH leaves preview.
Sources
- Microsoft, “Ignite 2026: Building the frontier ecosystem,” Satya Nadella, The Official Microsoft Blog, 17 November 2026
- Microsoft, “Microsoft Ignite 2026 Book of News,” 17 November 2026
i10x Perspective
These are vendor descriptions of a product launch and no outside tests are cited. Fully autonomous security agents raise accountability questions, which Microsoft addresses with playbooks and human control. More in our AI Security section.






