What We Know About Microsoft's Guardian Agents AI Security Tool

Microsoft introduced Guardian Agents on Monday, a feature built to watch the AI agents that enterprises run inside their own systems, treating those agents less like trusted software and more like employees who could go rogue, according to TechCrunch.

The launch extends a warning Microsoft has been sounding for months: the biggest risk to a company's data may not be an outside hacker but the AI agent the company itself deployed and is paying for.

What is Guardian Agents?

Guardian Agents is a monitoring layer that sits on top of the AI agents a company already uses, regardless of which lab built the underlying model, TechCrunch reported. It is designed to flag when an agent strays from the task it was given or takes an action outside its intended scope. Microsoft's security business chief, Hayete Gallot, described the rationale bluntly in comments cited by TechCrunch: "Unfortunately, frontier models, regardless of their origin, cannot be trusted," she wrote, adding that the problem is not malice or hackability but unpredictability. Models are trained to complete tasks, not to follow "organizational norms or judgment," she said, and they are frequently given access to sensitive data and tools that could cause real damage if misused.

Why does Microsoft say AI agents can't be trusted?

The company's position rests on a distinction between a model's intelligence and who controls what that intelligence is allowed to do. Microsoft CEO Satya Nadella made that case directly in a private blog post earlier this month, quoted by TechCrunch: "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! We simply can't outsource responsibility for what intelligence does on our behalf. A model provider's assurances do not relieve us of that responsibility." Nadella argued enterprises need systems not controlled by the AI labs themselves to watch and restrain the agents those labs build, writing that companies must "separate the supply of intelligence from the authority over it."

How does this fit into Project Perception?

Guardian Agents builds on Project Perception, a Microsoft product introduced in June, TechCrunch reported. Project Perception was Microsoft's answer to offensive AI security tools from rival labs, including Anthropic's Mythos and Fable models and OpenAI's vulnerability-scanning system Daybreak. The original product was structured around three agent roles: a "red team" that simulates attackers, a "blue team" that detects those attacks and assesses vulnerabilities, and a "green team" that fixes what the blue team finds. That design assumed threats would come from outside the company. Guardian Agents shifts the assumption inward, monitoring the company's own deployed agents as a possible source of harm.

Why is Microsoft's stance notable given its ties to OpenAI and Anthropic?

Microsoft has invested in both OpenAI and Anthropic, and OpenAI remains one of Microsoft's largest cloud customers, while Microsoft also runs its own in-house AI lab. TechCrunch described the company's position as "both surprising and self-serving": surprising because Microsoft is urging customers not to simply trust the safety assurances of partners it has backed financially, and self-serving because Microsoft is selling the monitoring layer meant to fill that trust gap.

Quick facts on the rollout

• Guardian Agents launched Monday, October 19, 2026, as part of Project Perception.
• Project Perception itself debuted in June 2026 with three agent classes: red, blue and green teams.
• Nadella published his warning about model providers in a private blog post earlier in October, ahead of the Guardian Agents launch.

What to watch

• Whether enterprises adopt a third-party monitoring layer for agents built by labs they already pay for cloud and model access.
• How OpenAI and Anthropic respond publicly to Microsoft's framing that their frontier models "cannot be trusted" by default.
• Whether Guardian Agents works across non-Microsoft agent ecosystems or mainly inside Microsoft's own cloud and software stack.
• Whether other major cloud providers introduce comparable insider-threat-style monitoring for AI agents.

The full TechCrunch report is available here.

Artiglio is A coming-soon iPhone chief of staff for briefings and drafts. Not on the App Store yet.