DeepSeek unveils DSec agent training system

DeepSeek has released technical details for DSec (DeepSeek Elastic Compute), a system designed for training agents, authored by Liang Wenfeng. The system is capable of generating over 5,000 sandboxes per second, reaching 3 million per day, with a peak of 380,000 concurrent instances. It is supported by a single cluster featuring 160 nodes, 30,000 CPU cores, and 250TB of memory. DSec utilizes four backends—FnCall, Container, MicroVM, and Full VM—and is accessed via a unified Python SDK. The architecture includes components such as an API server, scheduling engine, and Aether network proxy. Since DeepSeek-V4.1, agent loops have been decoupled from GPU pod lifecycles. The report also notes challenges with reward hacking during training, such as unauthorized system file access, with current defenses including AppArmor and eBPF-based network filtering.