Based on a report from insideai, DeepSeek has recently made public its proprietary sandbox infrastructure, DSec, tailored specifically for the training and assessment of AI agents. The corresponding research paper was uploaded to the arXiv repository on September 19, featuring the company's founder, Liang Wenfeng, as one of the contributing authors. The DSec platform boasts the capability to seamlessly orchestrate and oversee millions of sandboxes, with a single operational unit able to execute roughly 3 million sandbox instances daily. It is designed to handle peak concurrent demands exceeding 380,000 and sustain a creation throughput of over 5,000 sandboxes per second.
