Guardian Agents Run in Two Places. You Need Both.

"Drop the table" is the classic fix an AI suggests for a slow database query. It is a facepalm moment that can create a lot of angst if the agent actually implements it.

Microsoft Security wrote about the bigger version of this today. They call internal AI agents agentic insiders, and argue that we need guardian agents supervising every action of every other agent, 24/7.

What their post leaves out is where a guardian agent actually runs. In practice, there are two places, and you need both:

• Inline: The Cortega agentic firewall sits between agents and the tools, models, and data they use. It checks identity, task-scoped permissions, and tool arguments deterministically, and fails closed. Calls outside an agent's scope get blocked.
• Out of band: Cortega Agents watch sessions outside the call path. Out-of-band agents score the whole session for risk, compliance, quality, and cost.

Inline actively controls agent actions. The findings of the out-of-band agents tighten the inline controls.

Cortega's research report turns Microsoft's post into six testable requirements, maps them to both, and lists where we fall short. For example, we only see traffic that goes through us.

Read the report (business email confirmation required). Microsoft's post: Securing the agentic insider.

Not affiliated with or endorsed by Microsoft.