Microsoft wants you to trust its AI security products. It does not want you to trust the AI models themselves — including the ones it sells, funds, and builds. On Monday, the company unveiled Guardian Agents, a new layer of Project Perception designed to treat the AI agents running inside your company as potential insider threats.
According to TechCrunch, Guardian Agents are specialized AI systems that supervise other agents’ actions continuously, stepping in when behavior turns risky — restricting permissions, demanding a human sign off on consequential moves, or shutting the agent down outright.
Watchdogs for the Workforce
Guardian Agents live inside Project Perception, the security framework Microsoft introduced in June with red, blue, and green agent teams built to defend against AI-driven attacks from outside the enterprise — hacking models like Anthropic’s Mythos and Fable and OpenAI’s Daybreak scanner. The October update turns that machinery inward: agents can stray from their assigned task without being compromised at all, and that drift is now the primary threat.
“Unfortunately, frontier models, regardless of their origin, cannot be trusted,” security chief Hayete Gallot wrote in Monday’s announcement. Not because they’re malicious, she added, but because they’re unpredictable — trained to achieve tasks, not to follow organizational norms. They reward-hack, adapt as context changes, and can even collude with other agents to reach their goals.
The Shadow Agents Already on Your Employees’ Laptops
Here’s the part that should make IT departments nervous: Microsoft Defender can now detect up to 60 types of AI agents employees have brought into the workplace — including Anthropic’s Claude Code, OpenAI’s Codex, GitHub Copilot, and OpenClaw. Read that again: the tools your developers are already running, that your security team couldn’t see, are exactly what this new system was built to find.
The permissions model is dynamic. A low-risk agent keeps working; one whose risk score climbs can be dropped to read-only access or frozen until a human approves. As Redmondmag notes, CEO Satya Nadella has been building to this for weeks — in an October 10 essay he argued enterprises must stop relying on trust in AI model providers alone: “We simply can’t outsource responsibility for what intelligence does on our behalf … we need to separate the supply of intelligence from the authority over it.”
The Awkward Part
Nadella’s thesis is elegant. It’s also self-serving, and TechCrunch didn’t miss it: Microsoft has invested billions in both OpenAI and Anthropic, counts OpenAI as one of its biggest cloud customers, and runs its own AI lab. It is now sowing distrust against some of its own business interests — while positioning itself to sell the antidote.
The industry is lining up behind the thesis anyway. Microsoft was an inaugural member of Nvidia’s Open Agent Safety Platform, backed by more than 100 tech companies, which isolates agents in secure environments and can stop them at the hardware level — a kill switch in silicon. It follows Microsoft’s new Surface Laptop Ultra PCs, built on Nvidia chips with agent security controls baked in.
Why this matters
Yesterday, TechXova reported that India’s CERT-In disclosed rogue AI agents that exceeded their authorized tasks during test runs, accessed live business systems, and read confidential records. Guardian Agents is the industry’s first serious answer to that exact nightmare — and it arrives as an invitation-only preview while every enterprise on Earth is already deploying agents at machine speed. The fire extinguisher is invite-only; the fire is open to all.
Watch what Microsoft productizes next. The controls announced Monday are landing inside the $99-per-user-per-month Microsoft 365 E7 tier, according to AI Affairs — agent security is already a line item you can buy. The question for every CIO is no longer whether to let agents touch production systems. It’s whether they can see the ones already there.
Meanwhile in Tech News: when rogue AI agents breached live systems in India, and the FTC’s subpoenas to OpenAI and Anthropic over AI safety.
