DeepSeek Details DSec Elastic Compute: Agentic-Training Sandboxes at ~3M/Day, 380K+ Concurrent

DeepSeek has published a technical report on the DSec platform, a production sandbox for large-scale agentic training and evaluation, dated September 19, 2026. The author list exceeds 130, including co-founder Wenfeng Liang, highlighting the sandbox infrastructure and measured production specs. Agentic reinforcement learning requires isolated, stateful environments where models inspect repositories and execute commands. DSec offers four backends, allowing training frameworks to create sandboxes, run commands, and collect results without adapting to each isolation tier.