
Microsoft is expanding the reach of its AI-driven security suite, Project Perception, to confront risks that appear when autonomous software agents function inside corporate networks.
New focus on “agentic insiders”
Hayete Gallot, executive vice president of Microsoft Security, defines “agentic insiders” as autonomous entities that can retrieve data, employ tools and make decisions without constant human supervision. She cautions that while these programs can raise efficiency, they also open security gaps.
The platform first appeared in July and entered public preview on August 3. Its initial purpose was to help security teams counter sophisticated cyber threats through AI-powered automation.
Project Perception is built around three specialized AI groups. The “Red” units probe for potential weaknesses, the “Blue” units examine suspicious activity, and the “Green” units remediate problems and strengthen defenses.
To speed detection, the system gathers signals from identities, endpoints, cloud services, applications and additional sources, creating a full view of activity across the environment.
A recent ControlUp study highlights that sign-in problems are responsible for up to 53% of help-desk tickets at individual enterprises, showing how authentication failures can overwhelm support staff even before network issues arise, which often require longer resolution times.
The company now intends to apply the same capabilities to autonomous entities already present in enterprise networks. The suite will assign each program a unique identity, configure permission sets and continuously monitor its actions across systems.
At the same time, Gartner projects that global AI spending will climb to $2.67 trillion in 2026, with AI infrastructure comprising nearly $1.5 trillion—about 56% of the total—reflecting the massive investment needed to support platforms like Project Perception.
A newly introduced class called “Guardian Agents” will oversee other programs, stepping in when behavior turns risky. This may involve limiting privileges or blocking specific operations before any damage occurs.
Gallot wrote in the company blog that monitoring and securing those agents must become a continuous practice, emphasizing the need for real-time visibility into every autonomous action.
Executive perspective on AI risk
Satya Nadella’s essay published on October 10 framed powerful AI models as potential insider threats, especially when they can access confidential data or run critical processes.
He argued that organizations require controls that operate independently of the models themselves, such as limited permissions, detailed activity logs and mechanisms that let administrators halt a program’s actions.
“In other words, we need to separate the supply of intelligence from the authority over it,” Nadella wrote.
The firm acknowledges that observation alone will not solve the problem. Future tools must also explain why a model makes a particular choice, not merely record the outcome, to aid auditors and compliance teams.
Gartner also notes that for every dollar expected to be spent on generative AI models, more than $52 will be allocated to the supporting infrastructure.
Leave a Reply