The Paradigm Shift: From Data Protection to Agent Control
Corporate cybersecurity has undergone a fundamental turn. Traditionally, the focus centered on data protection and preventing unauthorized access by external actors. However, the proliferation of autonomous artificial intelligence systems has introduced a new category of risk: AI agents operating within enterprise infrastructure with a degree of autonomy.
Microsoft has responded to this challenge with Project Perception, a security architecture designed from scratch for the AI era. This system does not limit itself to detecting external threats; it redefines the concept of an 'insider threat' to include AI agents that are not under the direct control of development labs, such as OpenAI or Anthropic. The premise is clear: an AI agent that acts unpredictably or maliciously within the corporate network represents a risk equivalent to that of a malicious employee.
Project Perception Architecture: Perception, Reasoning, and Action
Project Perception is based on a multimodal architecture that combines signals, context, models, and specialized agents. Unlike traditional security systems that generate reactive alerts, this system is designed to perceive risk continuously, reason over large volumes of context, and act at machine speed.
The continuous risk assessment methodology relies on three types of specialized agents working in a closed loop. 'Red team' agents identify potential compromise paths before an attacker can exploit them. 'Blue team' agents investigate, reason from context, and determine what constitutes a real risk. 'Green team' agents execute corrective measures and reinforce defenses. This coordination allows a finding to become a proactive correction without the need for manual intervention at every step.
Roles of Agents in Project Perception
Distribution of functions among the three types of specialized agents in the system.
| Agent Type | Primary Function | Operative Objective |
|---|---|---|
| Red Team | Attack Simulation | Identify vulnerabilities and compromise paths before the attacker. |
| Blue Team | Investigation and Analysis | Discriminate real risks from noise and prioritize threats. |
| Green Team | Remediation and Reinforcement | Execute automatic corrections and strengthen defenses. |
Integration in Microsoft 365 E7: Specific Controls
The implementation of Project Perception integrates closely with the Microsoft 365 E7 suite, which incorporates specific tools to manage AI agent risk. The system uses the Microsoft Security Dashboard for AI, which provides a unified real-time view of the security posture of AI assets, including Microsoft Copilot Studio agents, Microsoft Foundry applications, and third-party agents.
Among the specific tools is the ability to detect and contain malicious or out-of-control agents. Microsoft Defender expands its coverage to monitor agent activity, while Microsoft Purview manages data classification and the prevention of confidential information loss through AI. Additionally, controls have been enabled to prevent the unauthorized use of AI applications (Shadow AI) and to avoid data leaks from enterprise browsers.
AI Agent Risk Assessment Flow
Step-by-step process of how the system evaluates and mitigates the risk of an AI agent in real time.
- 01Detection: The system identifies the AI agent in the AI asset inventory.
- 02Contextualization: Signals regarding the agent's identity, data, and behavior are collected.
- 03Evaluation: Blue team agents analyze the behavior to determine if it is anomalous or malicious.
- 04Decision: Action is prioritized based on the risk level and potential impact.
- 05Action: Green team agents execute the necessary containment or remediation.
Implications for Deploying Third-Party Agents
For companies deploying third-party AI agents, such as those developed by OpenAI or Anthropic, Project Perception introduces a new level of scrutiny. Microsoft emphasizes that frontier models cannot be trusted absolutely, making it essential to closely monitor the agents used by employees.
This implies that organizations must treat these agents as first-class identities, assigning them unique credentials and applying the principle of least privilege. Identity governance and least-privilege access are essential to limit the scope of agents and prevent misuse or attack. Additionally, continuous testing for message injection and jailbreaks is recommended to detect weaknesses before attackers do.
The Future of Agent Governance in the Corporate Cloud
The launch of Project Perception marks a milestone in the evolution of corporate security. Agent governance in the corporate cloud is no longer limited to permission configuration; it involves continuous supervision of the autonomous behavior of AI systems.
Microsoft insists that control remains in the hands of security teams. Organizations must define which actions can be executed autonomously, which require human approval, and how each decision is audited. This hybrid approach, combining machine speed with human judgment, will be key to balancing innovation with security in the AI era.
Open questions
- The specific metrics used by Project Perception to quantify the risk of an AI agent have not been detailed.
- It has not been specified how the evaluation of an AI agent technically differs from that of a human user in Microsoft 365 E7.
- The exact automatic actions the system can take when detecting an AI agent with anomalous behavior have not been defined, beyond general remediation.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction