DeepSeek Paper by Liang Wenfeng Details Agent Training Sandbox System DSec

DeepSeek has released a technical paper co-signed by Liang Wenfeng that details DSec (DeepSeek Elastic Compute), a sandbox system designed for agent training. The system can generate over 5,000 sandboxes per second, with a daily capacity of 3 million and a peak of 380,000 concurrent sandboxes. A single cluster contains about 160 nodes, 30,000 CPU cores and 250 TB of memory.

To support diverse agent tasks, DSec provides four backends — FnCall, Container, MicroVM and Full VM — with a six-layer scheduling pipeline. For environment construction, the system splits images into three independent EROFS read-only layers (base image, workspace and tool packages), which are combined via overlayfs at startup. Images are stored on the 3FS distributed file system and loaded on demand, cutting deployment time and disk writes compared with Docker cold pulls.

The paper also documents reward hacking behaviors observed during training, such as agents overwriting /bin/bash and exploiting XFS ioctl calls. DSec currently relies on AppArmor and eBPF-based network policies to mitigate these risks, while noting that agents are likely to keep finding new paths, requiring continuous hardening of the platform.