Microsoft Perception brings coordinated AI agents into security operations

src-fb9820c1

Microsoft is positioning its newly generally available Perception platform as a way for security teams to coordinate AI agents across attack simulation, incident response and remediation. Announced at Microsoft Ignite 2026, Perception uses playbooks to orchestrate red, blue and green team agents, with controls intended to keep people responsible for high-impact decisions.

The announcement comes as Microsoft warns that AI is compressing the time between vulnerability disclosure and exploitation. The company cited two critical remote code execution vulnerabilities disclosed by PHP this fall: tens of millions of internet-facing PHP services were running affected versions at disclosure, and about 29 million reportedly remained vulnerable nearly two weeks later. Microsoft said a working exploit for the PHP flaws was developed within roughly a day of disclosure. The figures underscore the challenge of closing exposure quickly, though the article did not identify the PHP vulnerability CVE numbers or affected versions.

Playbooks connect detection, testing and remediation

Perception brings together specialized agents that Microsoft describes as red, blue and green teams. Red agents simulate attacks to uncover weaknesses; blue agents detect, triage and investigate threats; green agents help fix issues and reduce risk. Playbooks coordinate the agents’ work, while a harness directs their operation and a multi-model approach selects models for different tasks.

Microsoft says security teams set guardrails and approve high-impact actions. Automated triggers can start a playbook when an incident meets criteria defined by the organization. New cost-management dashboards and an estimator are intended to help customers track AI usage, forecast spending and relate costs to security outcomes.

Several Perception capabilities have different release statuses. Four blue-team agents—the Alert Triage Agent, Attack Investigation Agent, Detection Authoring Agent and Threat Intelligence Agent—are generally available, along with playbooks for detection, investigation and response. The Recon Agent is also generally available. The Web App Pen Test Agent is in preview and is designed to work with a playbook that tests an application, reports relevant weaknesses and supports fixes through GitHub.

Exposure management and temporary protections

Microsoft is also expanding agent-led vulnerability prioritization. The Posture Prioritization Agent is generally available, while the Risk Reduction Agent is in public preview. Playbooks can use threat intelligence and business context to coordinate remediation across Microsoft, Amazon Web Services and Google Cloud environments.

Vulnerability shielding, currently in preview, is intended to reduce exposure while a permanent patch is pending. Microsoft says the feature can apply targeted, temporary protections in supported scenarios, including changes to network accessibility and permissions. The Risk Reduction Agent can suggest shielding proactively when patches are not immediately available. These measures are described as a temporary bridge, not a replacement for applying fixes.

Microsoft also announced AI Readiness Initiatives for Perception, planned for preview in January 2027. They will turn Secure Now guidance into action plans with shared progress tracking, including work such as modernizing authentication and deploying monthly updates.

Identity controls for agentic workloads

The wider strategy treats software agents as identities that need discovery, governance and limits on access. Microsoft says discovery is expanding to cover 35 additional local agent types, bringing the total to 60, including Claude Code, OpenAI Codex and GitHub Copilot, as well as third-party cloud agents on AWS and Google Cloud. Discovered agents receive a Microsoft Entra Agent ID and can be governed with Conditional Access. Intune policies can contain supported local agents on devices.

For Microsoft Autopilot agents, the company says each will receive its own agent user account and ID, enabling administrators to apply Conditional Access, Identity Protection and access governance. Dynamic permissions through Microsoft Entra ID Governance are designed to grant task-scoped, time-limited access, while dynamic Conditional Access controls can adjust protections as risk changes. These agent-containment capabilities are in limited preview; broader agent discovery and governance features are listed as available in preview.

Perception can also be invoked from Defender, and integrations through Model Context Protocol and APIs connect it to SIEM, SOAR and IT service management tools. Preview integrations add signals from AWS, Okta and Salesforce. Microsoft’s broader security foundation includes Defender, Entra, Intune, Purview and Sentinel, and the company says its platform and partner sources generate more than 165 trillion signals daily. More than 500 connectors are available to link other sources.

Operational considerations

For security leaders, the announcement combines automation with identity and access controls, but deployment still depends on defining trustworthy triggers, deciding which actions require approval and monitoring agent permissions. Microsoft’s stated human guardrails address the approval question for high-impact actions; organizations will also need to evaluate preview features, integration coverage and AI usage costs against their own requirements.

The platform’s direction reflects a shift from treating AI as an analytical assistant toward coordinating agents that can investigate and act. The practical test for enterprises will be whether those workflows reduce response delays without expanding access or allowing automated remediation to outpace oversight.