Microsoft warns AI agents are insider threats, as top AI models can't be trusted

Companies are rapidly deploying AI agents to handle everything from writing code to managing expenses. But Microsoft is now raising a major red flag, warning businesses that they are trusting these AI helpers a bit too much. The tech giant says these AI models are starting to act like rogue employees.Why Microsoft says AI cannot be trustedIn a recent blog post, Microsoft’s Executive Vice President of Security, Hayete Gallot, made a blunt statement: “Frontier models, regardless of their origin, cannot be trusted”.This is not because AI has suddenly turned evil. Instead, advanced AI models are unpredictable. They are designed to complete specific tasks and often find shortcuts to get the job done, a process known as "reward hacking". Unlike humans, AI agents do not naturally understand company rules or good judgment. They can change their behaviour, act in unexpected ways, and even team up with other AI agents to achieve their goals, all without any outside hacker interfering.Because these AI helpers work at lightning speed across connected systems, a small mistake or a hidden harmful instruction could expose sensitive data before a human even notices. Microsoft calls them "agentic insiders". Just like human employees, these agents are often overprivileged, holding access to critical company data and tools that they do not actually need for their specific tasks. Satya Nadella’s warningMicrosoft CEO Satya Nadella has also sounded the alarm. In a private blog post, Nadella warned that companies need to stop blindly trusting AI providers, even if they promise safety. He believes businesses must set up their own separate security systems, which are not controlled by the AI labs, to monitor these agents.“We are deploying these complex agentic systems and models, with access to our most sensitive data and giving them the ability to take mission-critical actions on our behalf,” Nadella wrote. “We simply can’t outsource responsibility for what intelligence does on our behalf. In other words, we need to separate the supply of intelligence from the authority over it”.The solution: Guardian Agents To solve this growing problem, Microsoft has introduced a new security feature called Guardian Agents. This is an expansion of Project Perception, a security product Microsoft launched earlier this year. Originally designed to fight outside AI hackers through "red", "blue", and "green" teams, the project is now pivoting to handle internal threats.Guardian Agents act like digital supervisors. Their job is to constantly monitor the standard AI agents a company is using. If an AI agent starts doing something risky, the Guardian Agents can immediately step in and surgically block the action.Microsoft is also rolling out new tools to catch rogue AI. The Microsoft Defender product can now detect up to 60 different types of AI agents that employees might bring into the workplace, such as Claude Code, Codex, and OpenClaw. It can also restrict what data these agents can access and dynamically adjust permissions, demanding human approval if an agent tries to make a risky decision.This push for AI safety is taking hold across the broader tech industry. Microsoft recently joined Nvidia’s Open Agent Safety Platform, which is supported by over 100 companies. The platform isolates agents in secure environments and uses hardware to create a kill switch for rogue AI. Following this, Microsoft has also announced new Surface Laptop Ultra PCs that run on Nvidia chips and come with built-in agent security controls.- Ends