Microsoft has introduced an AI-powered security platform to help organisations detect, investigate and respond to cyber threats. Called Microsoft Perception, it brings together agents that simulate attacks, examine suspicious activity and help fix vulnerabilities.
Announced at Microsoft Ignite on November 17, Perception appears to be Microsoft's answer to Anthropic's Mythos and OpenAI's Daybreak, reflecting a wider shift towards using AI to defend against increasingly sophisticated attacks.
Why Microsoft says security needs to change
Microsoft says the time needed to turn a software flaw into an attack has fallen sharply. It cited two critical remote code execution vulnerabilities disclosed in PHP this autumn. Such flaws can let attackers take control of servers over the internet.
At the time of disclosure, tens of millions of internet-facing PHP services were running affected versions. Nearly two weeks later, about 29 million services remained vulnerable, according to Microsoft. Researchers developed a working exploit for one flaw, known as PHP2Shell, within roughly a day of its disclosure.
Microsoft said more than 96,000 vulnerabilities were published under the Common Vulnerabilities and Exposures (CVE) programme in 2026, more than double the roughly 40,000 recorded in 2024. It also said only 26 per cent of vulnerabilities listed in the US Cybersecurity and Infrastructure Security Agency's Known Exploited Vulnerabilities catalogue were fixed in 2025.
Microsoft says AI helps attackers find weaknesses faster, enables more automated campaigns and creates risks when agents are manipulated, compromised or given excessive access.
How Microsoft Perception works
Perception is the agent layer of Microsoft's wider security architecture, which has three parts: signals and tools that collect information or take action, a context layer that organises information, and specialised agents that use it to carry out security work.
The first layer draws information from Microsoft products such as Defender, Entra, Intune, Purview and Sentinel, as well as networks, operating systems, browsers and cloud services. It also connects with other providers through more than 500 connectors. Microsoft noted that its own and partner sources generate more than 165 trillion signals a day.
The context layer combines these signals with information about an organisation's digital environment, giving agents a current picture of the systems they are protecting. The third layer consists of specialised agents, playbooks and AI models. Playbooks are coordinated steps for handling security tasks, with different models used for different jobs.
Red, blue and green agents have different roles
Perception uses three groups of agents. Red team agents take an attacker's perspective, simulating attacks and looking for weaknesses before criminals can exploit them. Blue team agents detect suspicious activity, sort alerts and investigate possible attacks. Green team agents work on fixing problems and reducing security risks. The groups work together to connect testing, detection and repair.
According to Microsoft, organisations remain in control. Security teams set guardrails and decide which actions need approval. Automated playbooks can start when an incident meets conditions set by the organisation, but teams must approve high-impact actions. Cost dashboards and an estimator will help customers monitor AI use and forecast spending.
Agents for threat detection and investigation
Four blue team agents are now generally available. The Alert Triage Agent assesses incoming alerts, while the Attack Investigation Agent investigates an incident. The Detection Authoring Agent helps create detection rules, and the Threat Intelligence Agent uses threat information to support security work.
Microsoft says the agents can move teams from alerts to investigations and improve rules for spotting similar activity, reducing manual work.
For proactive testing, the Recon Agent is generally available and helps identify and validate weaknesses. The Web App Pen Test Agent is in preview and can test web applications for vulnerabilities. A playbook can test an app, identify weaknesses and help drive fixes into GitHub code.
Finding and reducing vulnerabilities
Perception also includes tools to help organisations decide which weaknesses need attention first. The Posture Prioritisation Agent is generally available, while the Risk Reduction Agent is in public preview. Microsoft says they use threat intelligence and business context to prioritise fixes across Microsoft, Amazon Web Services and Google Cloud environments.
Vulnerability shielding, which is in preview, applies temporary protections to vulnerable systems when an immediate software update is unavailable. These can include changes to network access and permissions, reducing exposure while an organisation prepares a permanent fix.
According to the company, AI Readiness Initiatives in Perception are expected to enter preview in January 2027. They are intended to turn security guidance into trackable plans, including improving authentication and applying monthly updates.
Managing risks from AI agents
Microsoft's announcement also addresses risks from AI agents used inside organisations. These agents can work continuously, access business applications and data, and interact with other agents. Microsoft says they should be treated like digital insiders because they can make mistakes, be manipulated or have their access compromised.
Microsoft Agent 365 and Microsoft Entra capabilities can identify more types of local agents, including Claude Code, OpenAI Codex and GitHub Copilot, as well as some third-party cloud agents. The list is expanding by 35 types to 60.
Discovered agents can receive a Microsoft Entra Agent ID and be managed through access controls. Other measures include task-specific, time-limited permissions and controls that change as an agent's risk changes. Some of these capabilities remain in limited preview.
Microsoft also announced an AI Transparency Initiative. It says organisations should be able to see what an AI system was asked and authorised to do, understand the steps between a request and its outcome, and independently verify its actions. It is working with more than 40 organisations on standards and participating in an initiative led by the US National Institute of Standards and Technology.
How other companies are approaching AI cybersecurity
Microsoft is not the only technology company exploring AI for cybersecurity. Anthropic, OpenAI and Google have introduced models and tools to identify software vulnerabilities, investigate threats and fix security weaknesses.
Anthropic's Claude Mythos, for example, is designed for cybersecurity. The company has restricted access to the model through its Cyber Verification Programme because of concerns about its potential misuse.
Likewise, OpenAI has GPT-6 Astra and Daybreak. Google is also using AI for security through its Google Security Operations platform, which deploys specialised agents to analyse alerts and investigate threats. It has also developed Big Sleep to identify previously unknown software vulnerabilities and CodeMender to help find and fix software flaws.