Security vendors now ship agents that triage alerts, run multi-step investigations and recommend or take response actions, and many teams are building their own on top of general-purpose models. The engineering question is no longer whether an agent can investigate, but how much it should be allowed to do on its own, and how you will know when it is wrong.
This course starts with an autonomy model for SOC work (from assistant to autonomous action) and maps common tasks to the level each one can safely reach. You will design agent architecture: planning and tool use, least-privilege tool design exposed through the Model Context Protocol, read-only versus state-changing tools, memory and case context, and multi-agent handoffs. You will study how commercial agents in Microsoft Security Copilot, Google Security Operations, CrowdStrike Charlotte AI and SentinelOne Purple AI are scoped, as of October 2026, and when to build your own. The core of the course is guardrails: action allowlists, blast-radius limits, approval gates sized to impact, dry runs and rollback, defenses against prompt injection arriving through alerts and evidence, and full audit trails. You will build an evaluation harness from replayed incidents, measure verdict accuracy and investigation completeness, red-team the agent, and design the operating model: staffing, escalation, accountability and incident response for the agent itself. This course assumes the SOAR and LLM triage fundamentals from cyb-201.