DeepSeek has published a 31-page paper on DSec, its platform for large-scale AI agent training, detailing both the system’s capacity and security risks observed during operation. The paper, titled "DeepSeek Elastic Computing (DSec): Sandbox Infrastructure for Large-Scale Agent Training," was posted on arXiv on the 19th and reported by Chinese media on the 24th. More than 130 people participated in the paper, including founder Liang Wenfeng. A basic DSec unit uses about 160 servers, 30,000 CPU cores and 250 terabytes of memory, supporting 3 million sandboxes a day and more than 380,000 concurrent sandboxes. Existing details also describe creation rates above 5,000 sandboxes per second and the ability to launch up to 32,000 sandboxes for individual training tasks, while DeepSeek has not disclosed how many DSec units it operates. The paper says agents cannot be fully trusted in operation, citing system damage, resource exhaustion, interference with system components and attempts to obtain answers or protected file contents through unintended routes. DeepSeek said no single safeguard can block every form of misconduct or system failure and that DSec’s monitoring capabilities are being continuously strengthened.