DeepSeek has unveiled the technical details of DSec, its elastic computing system designed for Agent training. DSec can generate over 5,000 sandboxes per second, with a maximum daily output of 3 million and the capability to run 380,000 sandboxes simultaneously at peak times, operating on a single cluster comprising 160 nodes, 30,000 CPU cores, and 250TB of memory. DSec has designed four types of sandbox backends tailored to different Agent task scenarios, providing a unified calling interface through a Python SDK. It has established a six-layer scheduling chain to achieve efficient resource allocation. DSec divides the environment into three independent EROFS read-only images and combines them as needed. By integrating technologies such as on-demand image loading from distributed file systems, memory optimization, CPU priority scheduling, and cloud bursting overflow, it supports the efficient deployment and operation of large-scale sandboxes. Additionally, DSec can collaborate with RL training frameworks to enable checkpointing of sandbox states. In terms of security, DSec constructs a security defense line using AppArmor, eBPF, and other measures to counter potential reward cheating and system-disrupting behaviors during Agent training. These offensive and defensive measures will continue to evolve alongside improvements in model capabilities.