Artificial intelligence is changing not only how businesses operate, but also the nature of cyber threats. AI systems can analyse information, make decisions and perform actions without direct human involvement. This gives cybercriminals new opportunities to automate attacks, while organisations need security measures that can also operate faster and with greater context.
Microsoft is advancing a new AI agent-based cybersecurity concept in this direction. One of its most important components is Project Perception – a system designed to continuously detect risk, analyse threats and take protective action.
What does this mean for businesses? And why is AI security starting to require a fundamentally different approach from traditional infrastructure protection?
Key points at a glance
- Project Perception is a concept for an agent-based security system designed to continuously detect risk, analyse it and take protective action.
- The system uses three groups of specialised agents: Red Team, Blue Team and Green Team.
- Microsoft is developing a so-called cyber stack that combines data from across the IT environment, security context, AI models, agents and action execution mechanisms.
- One emerging threat is prompt injection, an attempt to manipulate an AI system using specially crafted instructions.
- Microsoft is also developing AI protection within Defender, Microsoft Agent 365, Entra and Purview.
- Project Perception entered public preview on 3 August 2026, according to Microsoft. Microsoft – Rethinking security for the age of AI
Why does AI security require a new approach?
AI is changing not only how users work. It is also changing the scale and pace of potential attacks.
A traditional cyberattack often requires time for the attacker to analyse the infrastructure, prepare a campaign, identify vulnerabilities and adapt subsequent stages of the attack. AI models can automate some of these tasks and enable large volumes of information to be analysed faster.
Microsoft notes that as autonomous systems evolve, the number of elements an organisation must secure also increases. This now applies not only to computers, servers and user accounts, but also to AI applications, agents, the data used by models and the relationships between these elements. Microsoft – Rethinking security for the age of AI
What is Microsoft Project Perception?
Project Perception is an agent-based security system being developed by Microsoft for environments where AI is part of an organisation’s infrastructure.
Its purpose is to bring together several elements:
- security signals,
- information about the organisation’s environment,
- AI models,
- specialised agents,
- mechanisms for executing protective actions.
In practice, this means moving away from individual tools operating independently towards a system that can combine information from multiple areas to create a unified view of the security situation.
Microsoft announced the public preview of Project Perception from 3 August 2026. Microsoft – Rethinking security for the age of AI
Red Team, Blue Team and Green Team – how do Project Perception agents work?
One of the most interesting aspects of Project Perception is the division of agents into three groups.
Red Team – identifying weaknesses
Red Team agents are intended to operate like a team conducting controlled attacks.
Their role is to identify potential paths to compromising the environment before a real attacker exploits them.
This may involve analysing:
- software vulnerabilities,
- misconfigurations,
- excessive permissions,
- dependencies between systems,
- potential paths an attacker could use to move through the infrastructure.
The key is to consider the chain of dependencies rather than a single alert in isolation.
For example, a vulnerability in an application may be of limited significance on its own. However, if the application has access to sensitive data, uses a privileged account and is connected to other systems, it becomes far more significant to the security of the entire organisation.
Blue Team – analysis and assessment of actual risk
Blue Team agents are intended to analyse signals relating to potential threats and combine them with the available context.
This is particularly important in large organisations, where security systems can generate enormous numbers of events.
An alert alone often does not indicate whether the organisation is actually at risk.
Relevant factors may include:
- user identity,
- the device from which the connection was made,
- the location and time of the event,
- the application in use,
- previous account activity,
- the user’s access to data,
- relationships between systems,
- previous incidents,
- information about current cybercriminal campaigns.
An agent can therefore analyse an event in a broader context rather than treating each signal as an isolated case.
Green Team – moving from analysis to action
The biggest difference emerges at the next stage.
Green Team agents are intended to carry out remediation actions and strengthen security controls.
Depending on the scenario, this may include:
- restricting access,
- changing the configuration,
- strengthening the security policy,
- responding to an identified vulnerability,
- securing a specific asset.
In this way, the three groups of agents form a closed security loop:
Red Team → identifies a weakness → Blue Team → assesses the risk → Green Team → takes action → the system analyses the environment again.
Microsoft describes this model as a continuous process of discovering, assessing and improving security. Microsoft – Rethinking security for the age of AI
The new cyber stack: what does AI-based security consist of?
Project Perception is part of a broader concept described by Microsoft as a cyber stack.
It can be simplified into several layers.
-
Signals – data and signals
The first layer provides information about what is happening in the environment.
Sources may include:
- devices,
- identities,
- applications,
- the cloud,
- data,
- AI systems,
- security solutions.
The more systems an organisation operates, the more important it becomes to connect them.
-
Security context – security context
Data alone is not enough.
The system needs to know what a given event means for a specific organisation.
For example, a sign-in from an unusual location may be normal behaviour for an employee travelling on business. The same event involving a privileged account may require an entirely different response.
Context therefore makes it possible to correlate information about assets, users, relationships, activity and risk.
Microsoft describes security context as a continuously updated view of the environment that enables agents to analyse situations more effectively. Microsoft – Rethinking security for the age of AI
-
Models – AI models
The next component is models responsible for reasoning and analysis.
Microsoft highlights a multi-model approach, where not every problem needs to be solved by a single, largest model.
A faster, less expensive model can be used for simpler tasks, while more complex analysis can be assigned to a model with stronger reasoning capabilities.
This approach can affect both performance and the cost of the overall system.
-
Agents – specialised agents
Agents perform specific tasks.
Rather than relying on a single “super-agent” responsible for everything, a system can use multiple agents specialised in specific functions.
-
Actuators – taking action
The final layer connects analysis with real-world response.
This is where a system decision can be translated into a specific protective action.
Microsoft emphasises that this layer is deeply integrated with Microsoft Security solutions. Microsoft – Rethinking security for the age of AI
Prompt injection – a new threat to AI systems
The growth of AI agents also creates new classes of threats.
One of them is prompt injection.
Put simply, it is an attempt to influence a model’s behaviour using carefully crafted instructions.
The issue becomes particularly significant when AI not only answers questions but also has access to data and tools and can take action.
Imagine an agent that analyses email.
An ordinary message may contain a hidden or explicit instruction directed at the model. If the agent treats it as a command rather than untrusted content, its behaviour could potentially be manipulated.
This is why AI security cannot be limited to protecting the model itself.
The following must also be secured:
model + data + context + tools + permissions + how actions are performed.
In July 2026, Microsoft announced that it was developing prompt injection protection in Microsoft Defender to identify and isolate email messages containing malicious instructions for AI systems before they are delivered to users. The feature was presented as a preview solution. Microsoft Security – What’s new in Microsoft Security: July 2026
An AI agent needs its own security layer
Introducing an AI agent into a company should not end with the question:
“Does the agent work?”
The following questions are equally important:
- What data can it access?
- Which applications can it use?
- What actions can it perform independently?
- What permissions does it have?
- Is its activity monitored?
- What happens when suspicious activity is detected?
- Can its access be revoked quickly?
This is why Microsoft is developing Microsoft Agent 365.
According to Microsoft, the new Defender capabilities are intended to combine security posture assessment and runtime protection for agents operating, among other environments, in Microsoft Foundry, Microsoft Copilot Studio and third-party agent management environments. Microsoft Security – What’s new in Microsoft Security: July 2026
In practice, this shifts the focus from user and device security towards the security of digital workers, meaning agents that perform tasks within the organisation.
Microsoft Defender: protecting the entire AI environment
The development of Project Perception does not mean that traditional security solutions are becoming less important.
Quite the opposite.
AI must be incorporated into the existing cybersecurity architecture.
Microsoft is expanding the capabilities of Defender, integrating information from different areas of the environment and using AI in SecOps processes. In its July update, Microsoft highlighted, among other developments, the evolution of Threat Intelligence Agent and the integration of additional analytics and automation features. Microsoft Security – What’s new in Microsoft Security: July 2026
This is important because AI security is not a separate project detached from the rest of the infrastructure.
An AI agent uses:
identities → devices → applications → data → networks → the cloud.
If any of these elements is inadequately secured, it can become part of a potential attack chain.
Can AI protect against AI?
This is one of the most important questions surrounding the development of Project Perception.
Microsoft’s answer is broadly affirmative: since AI enhances attackers’ capabilities, similar technologies should be used by security teams.
This does not mean, however, that people are no longer needed.
Microsoft emphasises the concept of human in control – AI systems should automate analysis and actions, but people should retain control over key security processes. Microsoft – Rethinking security for the age of AI
This is particularly important for high-impact actions.
Automatically blocking a suspicious process may be relatively safe. Automatically revoking an entire department’s access to a critical business system requires much greater oversight.
This is why the future of cybersecurity will probably not mean:
“AI instead of people”.
Rather:
“AI does more of the work; people make decisions where judgement and accountability are required”.
How can your organisation prepare for AI security?
AI implementation should begin with more than just selecting a model or application.
Preparing the right security model is equally important.
- Identify all AI use cases
Do not limit yourself to officially deployed tools.
Also review which AI services employees use independently.
- Identify agents and their permissions
Each agent should have a clearly defined scope of access.
The principle should be simple:
an agent receives only the permissions it genuinely needs.
- Secure your data
Define which data may be sent to AI models and external applications.
- Monitor activity
Knowing that an agent exists is not enough.
You must also know which data it uses, which applications it communicates with and what actions it performs.
- Prepare defences against prompt injection
Instructions from untrusted sources should be treated as potentially dangerous data, rather than automatically as commands for the model.
- Plan incident response
If an account, agent or application is compromised, the organisation should know how to quickly restrict access and stop the potential spread of the threat.
7 questions to ask before deploying AI agents
Before deploying autonomous agents in your organisation, it is worth answering several fundamental questions:
- What data will the agent have access to?
- What actions will it be able to perform without human involvement?
- What permissions will it receive?
- Will every agent activity be logged?
- How will we detect prompt injection or other manipulation attempts?
- What will we do if an agent is taken over?
- Can we quickly revoke the agent’s access to data and applications?
If the organisation does not have a clear answer to any of these questions, it should first strengthen its security foundations before increasing AI autonomy.
Summary
The AI era is also changing how organisations need to think about cybersecurity.
Organisations must protect not only users, devices, applications and data, but also AI models and autonomous agents, which can perform tasks and take actions within the organisation’s environment.
Project Perception is Microsoft’s response to this shift. Rather than a single tool, it offers a concept of a continuous, agent-based security system, combining data, context, AI models, specialised agents and automated response mechanisms.
At the same time, AI security requires a broader perspective than the model alone. The key elements are:
identity + permissions + data + applications + agents + monitoring + automated response.
How these elements are combined will determine whether AI becomes a source of new threats for the organisation or one of its most important tools for reducing them.
Would you like to learn more about Microsoft cybersecurity tools? Our team of experts will be happy to answer your questions and help you choose the best solution:
+44 (0) 20 3666 5846
You can also visit our website for more information about the products. Explore it and find out more: MICROSOFT
