Summary
- Microsoft Perception is generally available, coordinating specialised red, blue and green security agents through automated playbooks.
- Four blue team agents and the Posture Prioritization Agent are generally available, while Risk Reduction and vulnerability shielding remain in preview.
- New discovery and identity controls extend oversight to locally running AI agents, including coding assistants, with advanced governance capabilities still being tested.
Microsoft has made its Perception security platform generally available, combining specialised artificial intelligence agents to investigate threats, identify weaknesses and coordinate remediation across enterprise technology systems.
The launch, announced at Microsoft Ignite on 17 November, brings together AI-assisted security operations with additional controls designed to manage autonomous software operating inside organisations.
Perception coordinates agents responsible for different security activities through automated playbooks, while Microsoft is extending its Defender, Entra and Purview services to identify and restrict AI applications accessing enterprise information.
Although the core Perception platform is generally available, several newly introduced capabilities remain in preview. Organisations evaluating the products will therefore need to distinguish established functionality from features that are still being tested or rolled out to selected customers.
Security agents take on specialised responsibilities
Microsoft has organised Perception around three groups of security agents, reflecting established approaches to testing, defending and improving technology systems.
Red team agents examine environments from an attacker’s perspective, helping security teams identify and validate weaknesses before they are exploited. Blue team agents investigate alerts, examine suspicious activity and develop detection rules, while green team agents support the correction of identified problems.
The platform connects these activities through playbooks that can initiate a sequence of actions when specified conditions are met.
An investigation might begin with an alert, proceed through examination of related events and conclude with changes intended to improve detection or reduce exposure to similar attacks.
Microsoft says organisations can define the conditions under which automated processes operate and require human approval for actions carrying significant consequences.
That control is necessary because security responses can interrupt legitimate business activity. Blocking access, isolating equipment or changing system configurations may reduce an immediate threat while affecting employees and applications that depend on the resources involved.
The appropriate degree of automation will therefore vary according to the risk, the available evidence and whether a proposed action can be reversed safely.
Investigation capabilities enter general availability
Four blue team agents are generally available: Alert Triage, Attack Investigation, Detection Authoring and Threat Intelligence.
They are intended to help analysts assess alerts, investigate activity across systems, identify relevant threat information and develop detections for similar behaviour.
Microsoft has also made its Recon Agent generally available for security testing, while the Web App Pen Test Agent remains in preview.
Combining these capabilities could reduce the amount of manual coordination required when an incident moves from initial detection into investigation and remediation.
However, automated analysis depends on the completeness of the available security information. An agent may identify relationships between recorded events but cannot reliably assess activity that remains outside its monitoring scope.
Findings also require appropriate evaluation. Suspicious behaviour may have an authorised explanation, while a genuine attack can involve actions that resemble ordinary administration.
Microsoft is extending Perception integrations so customers can invoke playbooks within Defender and connect them to other security and IT management systems through APIs and Model Context Protocol interfaces.
Some integrations with external platforms and workflow tools remain in preview, meaning customers should verify support for their existing environments before treating the functionality as operationally established.
Vulnerability management moves towards automated action
Perception also includes capabilities designed to help teams prioritise vulnerabilities and coordinate remedial work.
The Posture Prioritization Agent is generally available, using information about exposure, exploitability and business context to help identify weaknesses requiring attention.
The Risk Reduction Agent remains in public preview, alongside new playbooks intended to connect prioritisation with the actions required to reduce exposure.
Microsoft has also introduced vulnerability shielding in preview, allowing supported systems to receive temporary protection where an appropriate software patch cannot be applied immediately.
Such protection might involve restricting particular network access or changing permissions while permanent corrective work is prepared.
These measures can reduce exposure without resolving the underlying software defect. Organisations still need to test and deploy suitable patches, establish whether temporary controls remain necessary and confirm that weaknesses have been corrected.
Security teams must also consider whether a proposed restriction would interrupt essential applications or create problems elsewhere in the environment.
Microsoft is expanding Codename MDASH, its AI-assisted code security scanner, to examine repositories of up to 5GB and support additional specialist security models.
The service remains in preview, although its integration with Perception could help connect security testing with investigations into weaknesses in application code.
Workplace AI agents become another source of exposure
AI agents increasingly operate on employee devices and through applications that were not necessarily designed around centralised enterprise management.
Developers may use coding assistants that interact with local files and remote services, while employees can install software capable of retrieving information and carrying out actions through connected applications.
These systems can inherit access to sensitive information or operate using permissions originally granted to a human user, creating difficulties when an organisation needs to determine which actions were performed by an employee and which were initiated by software.
Microsoft is expanding discovery to recognise 60 types of locally running AI agents, including Claude Code, OpenAI Codex and GitHub Copilot. The latest update adds 35 categories to its previous coverage.
The company is also extending Microsoft Entra Agent ID to provide supported agents with identities that can be managed separately from those of their human users.
Distinct identities can make it easier to monitor activity and apply controls according to an agent’s intended responsibilities. They may also reduce the need to grant software unrestricted access through an employee’s existing permissions.
Microsoft is developing dynamic permissions that can limit access according to the task and duration, while conditional access controls respond to changes in risk.
These advanced capabilities remain in limited preview, and the degree of support differs between agents and deployment environments.
Data protection follows autonomous activity
Microsoft Purview and Defender are gaining additional capabilities intended to monitor how AI agents interact with sensitive information and enterprise systems.
Data loss prevention policies can restrict selected information flows, while runtime security controls are being extended to identify potentially harmful actions during agent execution.
Such controls may help prevent confidential material being sent to unauthorised destinations or reduce the risk that malicious instructions cause an agent to misuse legitimate permissions.
However, effective protection depends on accurate information classification, appropriate policies and sufficient visibility of the applications involved.
Overly broad restrictions may interfere with legitimate work, while incomplete monitoring can leave some activities outside the intended safeguards.
Microsoft is also expanding Agent 365’s management capabilities, allowing organisations to discover and govern agents across different platforms, including selected third-party cloud services.
The company is developing Dynamic Agent Risk assessments intended to combine information about permissions, possible exposure and observed runtime behaviour. These capabilities are being introduced through preview programmes.
Centralised management may help organisations understand which agents have been deployed and who is responsible for them, but it cannot independently guarantee that the models driving those agents will behave appropriately.
Security and governance must therefore include application design, testing, access restrictions and procedures for handling unexpected outcomes.
Automation changes the operating model for defenders
Perception reflects an effort to move beyond AI systems that merely summarise alerts or recommend actions for analysts.
By coordinating specialised agents, Microsoft is seeking to automate more of the investigation and remediation process while retaining human approval for consequential decisions.
The commercial and operational value will depend on whether these systems can reduce investigation times, identify material threats accurately and improve recovery without introducing unacceptable disruption.
Organisations will also need to assess the cost of operating the agents, including model usage, integration, monitoring and the work required to review their results.
Microsoft is introducing usage dashboards and cost estimation capabilities intended to make that expenditure easier to assess.
Although the company offers a growing range of connected security products, customers may use different combinations of Microsoft and third-party services. The effectiveness of automated workflows will depend on which integrations are available and how reliably information moves between them.
The expansion of locally running AI agents introduces a related challenge: software that performs legitimate business tasks can also become an additional route through which information is exposed or actions are taken without sufficient oversight.
Microsoft Perception’s general availability establishes a wider operational platform for AI-assisted security work. The next evidence of its effectiveness will come from production deployments demonstrating whether automated investigations and remediation improve measurable security outcomes, while the company’s newer identity and runtime controls continue through their preview stages.












