Microsoft prepares defenses against unpredictable AI bots as trust erodes
Microsoft offers "police" for AI (collage: RBC-Ukraine)
Microsoft has introduced new tools to protect businesses from internal threats that autonomous AI agents might create, according to Microsoft.
Why are advanced AI models no longer trusted?
The launch of the new security system comes amid a series of warnings from Microsoft CEO Satya Nadella. According to him, businesses should stop blindly trusting AI developers' assurances about the safety of their products.
Nadella compared autonomous agents to unreliable insiders or employees who have access to critical company data but may act against its interests.
Microsoft's Head of Security Hayet Gallo emphasized that leading language models are not inherently malicious, but they are unpredictable.
They are optimized to achieve results at any cost and do not account for corporate norms or ethical boundaries, which often leads to the bypassing of security rules.
How do Guardian Agents and Project Perception work?
Microsoft's solution involves a clear separation between artificial intelligence providers and the systems that control them.
The Guardian Agents tool, part of the Project Perception initiative, employs specialized teams of protective algorithms (blue and green teams) that monitor the actions of other AI agents within the company's network in real time.
The system assesses the dynamic risk level of each action and can:
- Restrict AI bots' access to confidential databases and tools;
- Require human confirmation before executing critical operations;
- Automatically block the bot's activity if anomalous behavior is detected.
Furthermore, the updated Microsoft Defender has gained the ability to detect up to 60 types of third-party AI agents that employees might covertly use at work. This includes, in particular, Claude Code, Codex, and OpenClaw.
Contexts and new security standards
Microsoft's move reflects a broader trend in the IT industry. The tech giant has already joined the new Open Agent Safety Platform by Nvidia, which is supported by over 100 industry leaders.
This system isolates AI agents in a secure hardware-level environment and operates on the principle of an emergency kill switch.
The company is convinced that separating the provision of intelligence from its control is the only way to safely scale agent systems in business.
