Microsoft has introduced new security features designed to monitor and control artificial intelligence agents inside enterprises, reflecting growing concern that the systems companies deploy to perform complex tasks could become security risks themselves.
The measures, announced on Monday, include Guardian Agents, a feature intended to oversee AI agents used by organisations, and new capabilities in Microsoft Defender to detect workplace agents, restrict their access to data and systems, and adjust their permissions according to the tasks they perform. Some risky decisions can trigger requirements for human approval.
The launch follows repeated warnings from Microsoft chief executive Satya Nadella that companies should not assume AI agents are safe simply because their providers offer assurances. Nadella has argued that enterprises need independent systems to supervise agents, particularly when they have access to sensitive information and the authority to carry out critical operations.
“We simply can’t outsource responsibility for what intelligence does on our behalf,” Nadella wrote earlier this month in a private blog post. “A model provider’s assurances do not relieve us of that responsibility.” He added that companies needed to “separate the supply of intelligence from the authority over it”.
Microsoft’s new Guardian Agents feature forms part of Project Perception, a security initiative introduced in June. The original project divided defensive work among three categories of AI agents: a “red team” that simulated hackers and launched attacks, a “blue team” that detected those attacks and assessed vulnerabilities, and a “green team” tasked with fixing the weaknesses identified.
That approach focused on threats originating outside an organisation. Microsoft’s latest measures reflect a broader concern that the AI agents companies themselves deploy could pose risks similar to those associated with malicious or compromised employees, commonly described as insider threats.
Microsoft security business chief Hayete Gallot said frontier AI models could not be trusted to operate without appropriate oversight, regardless of where they originated. The problem, she wrote, was not necessarily that the models were malicious or easily hacked, but that their behaviour was unpredictable.
“They have been trained to achieve tasks, not follow ‘organizational norms or judgment’,” Gallot wrote. Their access to sensitive data and powerful tools could allow mistakes or inappropriate actions to cause significant harm.
The company’s position has implications for its own commercial relationships. Microsoft has invested in both OpenAI and Anthropic, maintains close ties with the two AI developers and counts OpenAI among its major cloud customers. Microsoft also operates its own AI research and development activities.
Its new security strategy nevertheless places emphasis on independent oversight of AI systems, including those supplied by companies with which Microsoft has business relationships. The company is positioning its own cloud and security products to provide that oversight, creating a commercial opportunity as businesses seek ways to manage the risks associated with increasingly capable agents.
In addition to Guardian Agents, Microsoft has introduced an agent-detection feature for Microsoft Defender that can identify up to 60 types of agents employees bring into the workplace, including Claude Code, Codex and OpenClaw. Other controls are designed to limit what agents can access and change their permissions dynamically according to the task being performed.
The approach is gaining support across the technology industry. Microsoft was an inaugural member of Nvidia’s Open Agent Safety Platform, an initiative backed by more than 100 other technology companies. The platform is designed to isolate agents in secure environments and use hardware-based monitoring to detect and stop potentially harmful behaviour.
Microsoft has also announced Surface Laptop Ultra PCs built for AI agents, using Nvidia chips and incorporating agent security controls into the hardware and Windows 11 environment.
