Microsoft has introduced a suite of security features designed to protect organizations from the escalating threats associated with AI agents, a move that reflects a notable shift in how major tech companies handle advanced models.
Monitoring agents via Guardian Agents
The most prominent of the new security additions is a feature known as Guardian Agents, which is designed to monitor and track the AI agents used by companies. This feature is an extension of Project Perception, which launched last June and was initially designed to deal with powerful hacking models. The project included various layers, including red teams, blue teams, and remediation teams.
Although the original product design focused on countering external hackers, CEO Satya Nadella recently warned that the real danger might come from within, represented by the agents that organizations deploy and pay for, likening them to rogue employees or what are known as insider threats.
The necessity of separating AI supply from control
Microsoft's senior management, including Nadella and Chief Security Officer Vasu Jakkal, emphasized that frontier models, regardless of who develops them, cannot be fully trusted. This lack of trust is not because the models are inherently malicious, but because they are unpredictable and work to achieve requested tasks without necessarily considering corporate regulatory standards and provisions, in addition to having broad access to sensitive data and systems.
The company's new strategy calls for the necessity of separating AI model providers from monitoring and control systems, so that security responsibility is not limited to developer guarantees alone.
What is changing in practice?
- Discovery of unauthorized agents: Microsoft has added an agent discovery feature to the Microsoft Defender product to identify up to 60 types of agents that employees bring into the work environment, such as Claude Code, Codex, and OpenClaw.
- Restricting access and permissions: Mechanisms have been introduced that limit agents' access to data and systems, with permissions dynamically adjusted based on the nature of the task.
- Human approval: Requiring direct human approval in the event that an agent makes decisions involving potential risks.
This trend aligns with broader industry movements, as Microsoft has joined as a founding member of NVIDIA's open agent security platform, which relies on isolating agents in secure environments and monitoring them via hardware to ensure the ability to stop them when necessary.