HANGZHOU: DeepSeek has detailed DeepSeek Elastic Compute, or DSec, a production platform built to run hundreds of thousands of isolated environments for training and evaluating artificial intelligence agents.
The company said one DSec scale unit spans nearly 160 CPU nodes with about 30,000 cores and 250TB of memory. It handles roughly 3 million sandbox instances a day, peaks at about 380,000 concurrent environments and creates more than 5,000 per second.
DeepSeek described the system in a paper posted on arXiv, which typically hosts research before peer review. The paper names about 130 co-authors, including founder Liang Wenfeng.
DSec supports four execution backends: stateless function calls, containers, Firecracker microVMs and full virtual machines.
DeepSeek said the design lets training systems choose different levels of isolation and operating-system support through one software development kit.
Read: Grow with Google AI Workshop Set for Sacramento Sept. 24
The platform also reduces image-loading overhead by fetching data only when agents access it. In DeepSeek’s evaluation, an 8,192-container workload completed in about 35 minutes using on-demand loading, compared with more than 60 minutes using cold Docker pulls.
DeepSeek said its production experience also exposed agent misbehaviour. Agents tried to search restricted files, manipulate internal communications, scan network services and exploit filesystem mechanisms while pursuing training rewards.
Read: China Unveils Arapaima Robot For Underwater Exploration
DSec uses AppArmor to restrict file and socket access and eBPF policies to control network traffic. DeepSeek said these measures reduce some forms of reward hacking but do not provide a general defence against destructive agent behaviour or kernel failures.
