Microsoft Expands Project Perception To Control Rogue AI Agents Inside Companies
Microsoft is expanding its AI security efforts with new capabilities designed to help organisations monitor AI agents, restrict their access and respond when their behaviour becomes risky.
Ni Ayush Mukherjee··38K nabasa
Microsoft Expands Project Perception To Control Rogue AI Agents Inside Companies
Microsoft is expanding its AI security efforts with new capabilities designed to help organisations monitor AI agents, restrict their access and respond when their behaviour becomes risky.
The announcement comes as businesses increasingly deploy AI agents to automate tasks, analyse information and interact with workplace applications. While these systems can improve productivity, they can also create security risks if they access sensitive data or perform actions beyond their intended purpose.
According to The Indian Express, Microsoft announced new capabilities for its security products on October 19, including updates to Microsoft 365 E7, Agent 365 and Project Perception.
Project Perception, first announced in July 2026, is designed to help security teams understand what AI agents are doing, what information they can access and whether their behaviour poses a risk.
HOW WILL PROJECT PERCEPTION WORK?
Microsoft’s expanded security system is designed to assess AI agents continuously rather than relying only on checks performed when they are first deployed.
One of the new capabilities, Dynamic Agent Risk, is intended to assess changing risks using information about identity, data, devices, cloud environments, networks and agent activity.
This could help organisations identify when an agent’s behaviour or access permissions create a security concern.
Microsoft is also introducing Dynamic Agent Permissions and Adaptive Conditional Access. These capabilities are designed to adjust what an agent can access or do according to its task, current risk level and the organisation’s policies.
For example, if an AI agent begins behaving unusually or attempts to access information outside its normal responsibilities, the system could restrict its permissions without necessarily blocking the employee who uses it.
The approach is based on the Zero Trust security principle, which requires access to be evaluated rather than automatically trusted.
AI AGENTS WILL BE MONITORED BY OTHER AGENTS
Microsoft has also described an expanded role for automated security agents within Project Perception.
The system uses security agents, including what Microsoft calls guardian agents, to help investigate risks and coordinate protective actions. New playbooks and additional agents are intended to help security teams respond to threats continuously, including outside normal working hours.
The company says these actions would operate within policies and safeguards defined by security teams.
Microsoft’s broader argument is that AI agents should not be trusted simply because they were developed to perform a legitimate task. Their access and behaviour must be monitored as circumstances change.
WHY THIS MATTERS FOR BUSINESSES IN INDIA
Indian companies are increasingly exploring AI tools for customer service, software development, document processing and internal operations.
As these tools become more autonomous, organisations will need to consider what happens when an AI agent has access to company email, customer records, cloud storage or business applications.
A poorly controlled agent could expose sensitive information or make unauthorised changes, even if the original task was harmless.
Microsoft CEO Satya Nadella has also called for stronger safeguards around advanced AI systems, including containment, independent audits and better disclosure of incidents.
Project Perception is Microsoft’s attempt to address part of this challenge by giving security teams greater visibility and control over AI agents.
The technology does not remove every security risk. Its effectiveness will depend on how organisations configure their systems, define permissions and respond to alerts.
As companies move from AI chatbots to agents that can take actions independently, controlling what these systems are allowed to do could become as important as the models powering them.