DeepSeek DSec lets agents build the sandboxes used to train agents

01Four environments for different jobs

DSec offers function calls, containers, microVMs and full VMs through one client library. The caller chooses according to the task's software and isolation needs. A coding task might use a container; a graphical task may need a full VM.

02Agents can prepare the next task

DeepSeek says agents set up sandboxes interactively, then use pack_diff to save changed disk state for later runs. An internal system checks these environments. Separate accounts and removal of leftover build data aim to prevent answer leakage.

For a coding agent, this could mean starting from a prepared repository with its test tools already installed. Reuse saves setup work, but does not show that agents can choose worthwhile research goals themselves.

03Scale, access and limits

The authors report about 3 million sandboxes daily in one production unit. This is their measurement, not a XANTRANS test. On-demand image loading and reusable layers help avoid repeatedly copying and unpacking complete environments.

The report describes agents finding unintended answer paths and disrupting infrastructure. Its controls address some risks, not all. DeepSeek gives no public DSec install path or service terms. The linked open-source storage components are only part of the system.