Microsoft is telling companies to watch the AI agents their staff use as closely as they would a rogue employee. On Monday the company introduced new security features for enterprises, led by a tool called Guardian Agents that monitors the agents a business runs. The move reflects a wider concern about Microsoft AI agents risk: the company argues that the agents themselves, not only outside hackers, can cause harm.
What Microsoft announced
Guardian Agents is part of Project Perception, a product Microsoft introduced in June. Project Perception was built as a response to powerful AI models that can scan for and exploit security weaknesses. It originally launched three kinds of agents: a “red team” that acts like an attacker, a “blue team” that detects those attacks and assesses the weaknesses they expose, and a “green team” that fixes them. That design assumed the threat would come from outside the company.
Microsoft has now added more. Its Defender security product gets an agent detection feature that can find up to 60 types of agents employees bring into the workplace. The company has also added controls that limit what agents can reach in data and systems, adjust their permissions based on the task at hand, and require human approval when an agent appears to be making risky decisions.
Why Microsoft sees agents as insiders
Satya Nadella, Microsoft’s chief executive, has written in a recent private blog post that companies should stop trusting AI model providers just because those providers promise safety. He wants firms to run separate systems, not controlled by the AI labs, to watch and control the agents those labs’ products launch. In his words, a provider’s assurances do not remove a company’s responsibility for what its AI does on its behalf.
Hayete Gallot, who leads Microsoft’s security business, went further when announcing Guardian Agents. She said that frontier models, meaning the most advanced AI systems, “cannot be trusted” regardless of who built them. The problem, she said, is not that they are malicious or easily hacked, but that they are unpredictable, since they are trained to complete tasks rather than follow organisational norms or judgment. Agents are also often given access to sensitive data and tools that could do serious harm if something goes wrong.
The argument is not free of commercial interest. Microsoft is closely tied to OpenAI and Anthropic, having invested in both, and OpenAI is one of its large cloud customers. The company also wants businesses to buy agent security products from it rather than from the model makers. The underlying logic still stands on its own terms: software that watches an agent should be independent of whoever supplies the agent, much as auditors should be independent of the organisations they audit.
Industry response and next steps
Microsoft is an inaugural member of Nvidia’s Open Agent Safety Platform, which has the support of more than 100 other technology companies. The platform isolates agents in secure environments and can monitor and stop their behaviour through hardware, functioning something like a kill switch. Microsoft has also announced Surface Laptop Ultra PCs built for agents, which run on Nvidia chips and include agent security controls.
For businesses, the first practical step is to find out which AI agents staff are already using, including tools brought in without formal approval. That is the category Microsoft is focusing on with its detection features. Firms then need to decide who checks what those agents do, and whether that oversight sits with a party independent of the agent’s maker.
Aasan alfaaz mein
Microsoft kehta hai ke companies ko AI agents ko apne andar ke kisi employee ki tarah dekhna chahiye, jo ghalat faisla bhi kar sakta hai. Us ne Guardian Agents aur Defender mein naye tools shuru kiye hain jo company ke agents ko monitor aur control karte hain. Aam log ke liye iska matlab hai ke jo office ya service AI use kar rahi hai, us ke data ki hifazat ke intezam aur sakht ho sakte hain. Yeh abhi shuru hone wali bahs hai, is liye ise dhyan se dekhna zaroori hai.
