Skip to content
CyberSimStart training

Theory · 10 min

Monitoring AI agents in the SOC

Write a first hunt for an AI agent doing something it should not.

An agent leaves the same telemetry as any program: process creation, file access, network connections, cloud API calls. On an endpoint the agent is the parent process, so its children are where to look.

MITRE’s detection strategy for T1700 describes the pattern to hunt: an AI-associated process accesses data outside its normal scope, then, within a short window, runs something that can transfer data and connects to a destination outside the approved AI endpoints.

Useful signals: an agent reading credential files (.ssh, .aws, browser profiles), starting shells or package runners (npx, pip) with packages nobody pinned, and connecting to new domains or webhooks.

Agent and MCP logs hold what endpoint logs cannot: the prompt, the tool calls and the tool results. Make sure they are collected before you need them.

T1700T1552.004

Your turn: build a query clause by clause. Build a Defender advanced hunting query: processes started by the coding agent devagent.exe that touch an .ssh folder.

Do the interactive check Free, in the browser.

Sources

Part of