Evidence-backed analysis
Microsoft spans assistive triage, coordinated defense, and vulnerability research
Security Copilot agents, Project Perception, and MDASH solve different classes of work.
Human reviewed 2026-08-27 · ASI-1.0 ·5 evidence records
Four systems, not one maturity story
Security Copilot Agents is a broad agent framework. Security Alert Triage Agent performs a narrower Defender classification workflow. Project Perception coordinates red, blue, and green agents in a limited public preview. MDASH is a vulnerability-discovery harness used internally and tested with a small private-preview customer set.
What actually differs?
The triage agent can classify supported alerts and is managed through administrator permissions and feedback. Project Perception coordinates multi-agent playbooks and requests approval for critical decisions. MDASH discovers and validates software vulnerabilities but does not establish autonomous patch deployment.
Availability matters
Project Perception remains invitation-only limited public preview in the reviewed source. MDASH remains limited private preview for external customers. A buyer cannot safely infer broad commercial availability from internal Microsoft use or benchmark headlines.
Evidence limit
Microsoft publishes substantial architecture and workflow documentation. The corpus does not contain qualifying independent production-effectiveness evidence for these systems. Benchmark scores reported by Microsoft remain vendor-reported unless the benchmark operator or an independent evaluator publishes the scoped result.
Governed evidence
Sources supporting this analysis
VENDOR DOCUMENTATION EV-MS-AGENTS-AUTOMATE
Microsoft Security Copilot agents automate repetitive tasks and reduce manual workloads. They respond to user requests and system events.
Canonical source: Microsoft →VENDOR DOCUMENTATION EV-MS-TRIAGE-CAPABILITY
The Phishing Triage Agent uses large language model based analysis to assess reported emails, determine intent, and classify each submission as a real threat or a false positive.
Canonical source: Microsoft →VENDOR DOCUMENTATION EV-PERCEPTION-MULTIAGENT
Project Perception is a new multi-agent security system designed for the realities of AI.
Canonical source: Microsoft →VENDOR DOCUMENTATION EV-PERCEPTION-GATE
Throughout the workflow, agents surface recommendations and request approvals when needed, keeping security teams in control of critical decisions.
Canonical source: Microsoft →VENDOR CLAIM EV-MDASH-AGENTS
The harness orchestrates more than 100 specialized AI agents across an ensemble of frontier and distilled models to discover, debate, and prove exploitable bugs end-to-end.
Canonical source: Microsoft →