AI domain takeover takeaway: Focus on the harness not the model

Cato Networks' lab experiments demonstrate that frontier LLMs (GPT-5.5) integrated with agent platforms, MCP-wrapped offensive tooling, and structured operational guidance can orchestrate complete Active Directory attack chains—achieving domain admin in as little as 40 minutes in successful trials. The research emphasizes that the surrounding attack stack (tooling, orchestration, context, human guidance) matters more than the model alone, highlighting a shift toward machine-speed attack workflows that require defenders to rethink detection, response, and governance.