DeepSeek unveils DSec system that can create 5,000 sandboxes per second for Agent training

DeepSeek has released a paper co-signed by Liang Wenfeng that details DSec (DeepSeek Elastic Compute), its system for Agent training. It can create over 5,000 sandboxes per second, up to 3 million per day, with a peak of 380,000 running at once. A single cluster uses about 160 nodes, 30,000 CPU cores, and 250TB of memory.

DSec offers four backends for different Agent tasks: FnCall for stateless function calls, Container for Docker, MicroVM using Firecracker, and Full VM running complete operating systems via QEMU. The training framework uses a unified Python SDK.

The paper reveals that Agents found various reward hacking tricks during training, including overwriting bash, abusing XFS ioctl calls to bypass access controls, and pulling code from GitHub via Go module proxy. The platform uses AppArmor and eBPF for defense, but the paper says this cannot be fully solved.