CrowdStrike builds seven agent-containment layers

CrowdStrike has implemented 7 independent control layers for autonomous agents and says none of its offensive agents has escaped the intended sandbox boundary in testing to date. Its agent-containment architecture nests network, virtual-machine, container, and process isolation beneath separate MCP and command gates, then caps tool calls, concurrency, runtime, compute, and spend; sensitive actions pause for human approval and time out closed. The design turns runtime enforcement inside the agent harness into an operator checklist: treat the model as untrusted code, keep credentials outside its process, and make each control survive the failure of another.