Microsoft launches Guardian Agents, warns AI agents can act as insider threats

Microsoft said on Monday, 19 October 2026, that it is treating AI agents deployed inside companies as a new category of insider threat, launching a security feature called Guardian Agents as part of its Project Perception platform, TechCrunch reported.

What Microsoft announced

• Guardian Agents is an addition to Project Perception, which Microsoft introduced in June with three classes of agents — a "red team" that attacks like a hacker, a "blue team" that detects it, and a "green team" that fixes vulnerabilities — according to TechCrunch.
• Microsoft's security business chief, Hayete Gallot, said when announcing the feature: "Unfortunately, frontier models, regardless of their origin, cannot be trusted," adding this is because the models are unpredictable rather than malicious, TechCrunch reported.
• Chief executive Satya Nadella said: "We need to separate the supply of intelligence from the authority over it," TechCrunch reported.
• Microsoft's Defender product has added an agent-detection feature that covers up to 60 types of workplace agents, including Claude Code, OpenAI Codex and GitHub Copilot, after the company expanded discovery to 35 additional agent types, according to Microsoft's security blog.
• The blog also introduced Dynamic Permissions and Dynamic Conditional Access, which can narrow, time-limit or revoke an agent's access if its behaviour changes mid-task, Microsoft said.
• Microsoft was a founding member of Nvidia's new Open Agent Safety Platform, which TechCrunch said has the backing of more than 100 other technology companies, isolates agents in secure environments and can monitor and stop their behaviour through hardware.

Why it matters

Microsoft's move formalises a shift from treating AI agents purely as productivity tools to managing them as potential security risks that need independent monitoring, containment and revocable access — separate from the AI labs that build the underlying models, even where Microsoft has invested in those same labs. For companies and public bodies across the Caspian region now piloting agentic AI in customer service, finance or coding, the episode is a reminder that deploying an agent built by one vendor does not come with built-in oversight of what that agent is permitted to touch, a gap enterprises will increasingly be expected to close themselves.