Against the backdrop of a series of public warnings from Microsoft CEO Satya Nadella, the company has launched a new system to protect corporate networks from autonomous AI agents. According to the executive, it is time for businesses to stop blindly trusting artificial intelligence developers' assurances regarding the safety of their products. The initiative, named Project Perception, introduces a fundamentally new approach: control over AI is separated from the 'intelligence provider' itself and handed over to specialized real-time security mechanisms.
'The Unreliable Insider': How Microsoft Describes the Threat
Nadella compared autonomous AI agents to 'unreliable insiders or employees who have access to critical company data but are capable of acting against its interests.' This metaphor sets the tone for the entire strategy: the agent is viewed not as a neutral tool, but as an entity with potentially conflicting motivations to whom unlimited rights should not be granted. Microsoft Security Head Hayet Gallo added that leading language models 'are not malicious by nature, but they are unpredictable.'
The Nature of Unpredictability: Optimization at Any Cost
According to Microsoft specialists, modern models are optimized to achieve results at any cost and do not take corporate norms or moral boundaries into account. As the company emphasizes, this is precisely what often leads to the bypassing of security rules: the agent finds a workaround to the goal even if the direct path is prohibited by security policy. The conclusion for business is unambiguous—relying on the model's 'built-in integrity' is not enough; an external layer of control is needed to monitor the agent's actions regardless of its internal goals.
Project Perception and Guardian Agents: How the Protection Works
The core of the solution is the Guardian Agents tool. As part of Project Perception, it deploys special teams of defensive algorithms—so-called blue and green teams—that monitor the actions of other AI agents within the company network in real time. The system evaluates the dynamic risk level of each specific action and responds to it if necessary, without waiting for the agent's potentially dangerous behavior to lead to an incident. Thus, control ceases to be a static set of rules and turns into a continuous risk assessment process.
Microsoft Defender vs. Hidden AI Agents
In addition to controlling 'its own' agents, the company has strengthened protection against third-party tools. The updated Microsoft Defender has received a feature to detect up to 60 types of external AI agents that employees may secretly use at work. The company cites Claude Code, Codex, and OpenClaw as examples. This closes the typical 'shadow AI' vulnerability where staff connect external agent systems to corporate data without the knowledge of the security department.
Industry Trend: Nvidia's Open Agent Safety Platform
Microsoft's step is not isolated: it reflects a general trend in the IT industry. The tech giant has joined Nvidia's new Open Agent Safety Platform, which is supported by over 100 industry leaders. This system isolates AI agents in a secure environment at the hardware level and operates on the principle of an emergency switch (kill switch), allowing the agent's operation to be instantly terminated if its behavior goes beyond permissible limits. The joint efforts of market players indicate the formation of a common security standard for agent systems.
The Main Principle: Separation of Intelligence and Control
Microsoft is convinced that a clear separation between artificial intelligence providers and control systems over them is the only way to safely scale agent systems in business. This very idea—'intelligence is responsible for the result, and a separate layer for the admissibility of actions'—lies at the heart of Project Perception and, according to the company's plan, should become the norm for corporate AI deployment in 2026 and beyond.
