Microsoft Perception AI Security Platform Goes Live

Microsoft has taken its agentic security platform out of preview, making Microsoft Perception generally available at its Ignite 2026 conference. The system coordinates teams of AI agents that handle security work inside Microsoft Defender, from sorting alerts to probing for weaknesses and fixing them.
Microsoft first introduced the project as Project Perception in July, and it has been in public preview inside Defender since August. At Ignite, the company dropped the project name, declared the platform generally available, and added new capabilities that widen what the agents can do across a security operation.
The announcement comes as attackers use AI to move faster. Microsoft said a phishing campaign powered by a phishing-as-a-service kit took about fourteen days to run its course a year ago, while comparable activity observed more recently unfolded in just 85 minutes and spread from 10 targeted users to 3,000. Perception is Microsoft's answer to that speed: defence carried out by agents working in parallel, rather than one alert at a time.
Red, blue and green agents
Perception splits its agents into red, blue and green teams. Red agents play the attacker and look for ways into a system. Blue agents handle detection, alert triage and investigations. Green agents focus on fixing problems and reducing risk. Playbooks orchestrate the right agents around a specific security outcome, and security teams set the priorities and approve critical actions.
Four blue-team agents are now generally available, covering alert triage, attack investigation, detection authoring and threat intelligence. Microsoft said its Alert Triage Agent can handle about 80 percent of alerts in a typical Defender queue, based on an analysis of about 8 million alerts across its security products. The agent reaches a verdict on each alert and explains its reasoning so analysts can review the evidence and act.
Once triage narrows the queue to confirmed threats, the Attack Investigation Agent goes deeper. It reasons across endpoints, identities, email, Active Directory and cloud resources, reconstructs the attack timeline, identifies what was compromised, and recommends a response such as containing an account or isolating a device. The Threat Intelligence Agent can turn external threat articles into structured findings, and the Detection Authoring Agent can translate that intelligence into detection rules tested against historical data before they are switched on.
The agents are no longer confined to Microsoft's own signals. The Attack Investigation, Detection Authoring and Triage Agents can now reason across signals from third-party vendors, including Okta, AWS and Salesforce, with support for more vendors to follow.
Pen tests and exposure management
On the red-team side, the Recon Agent is generally available, and a Web App Pen Test Agent has entered preview. The pen-test agent examines an authorised web application from an attacker's perspective, testing authentication, application programming interfaces, injection flaws and business logic, and reports which weaknesses are actually exploitable, with evidence and proofs of concept. In an independent benchmark run with Incalmo, Microsoft said the agent achieved a 31.0 percent average task solve rate, about twice OpenAI Codex's 15.2 percent, while staying within specified testing boundaries in 91 percent of safety-constrained tests.
Two new green-team playbooks work across Microsoft, AWS and Google Cloud. The Posture Prioritization Agent, which is generally available, helps teams decide which posture risks to tackle first. The Risk Reduction Agent, in public preview, turns a prioritised exposure into an action plan and can recommend temporary shielding, such as web application firewall rule changes, when a permanent fix is not yet available.
Agents under watch too
Microsoft also addressed the risk posed by AI agents themselves, which it treats as a new kind of insider with broad access to data and tools. In preview, Agent 365 and the Entra Suite can detect 60 types of local AI agents, including Claude Code, OpenAI Codex and GitHub Copilot, while Intune policies can keep supported agents on the device.
Perception can also be linked to the tools security teams already use. Through application programming interfaces and Model Context Protocol support, defenders can start playbooks from IT service management, security information and event management, or security orchestration systems, and bring agent sessions and artefacts back into those workflows.
Ales Holecek, president and chief architect of Microsoft Security, said teams set the guardrails and sign off on high-impact actions, with automated triggers launching playbooks when an incident meets criteria the team has set. The message from Microsoft is that agents do the volume work, while people keep control of the decisions that matter.