DeepSeek has detailed its DeepSeek DSec (DeepSeek Elastic Compute) platform in a technical paper, outlining the sandbox system the company uses to train AI agents at scale. The paper on arXiv features roughly 130 authors, including founder Liang Wenfeng, and was highlighted by outlets like Bloomberg this Wednesday (23).
What is a sandbox here?
A sandbox is an isolated environment: the agent (an AI model that interacts with the computer rather than just generating text) receives tasks, executes commands, and gets evaluated. If it messes up files or tries to cheat on the training reward, the damage is contained. DeepSeek DSec offers four levels of isolation—from function calling to full virtual machines—via a single SDK.
The numbers
A production unit consists of roughly 160 nodes, 30,000 CPU cores, and 250 TB of memory. According to the paper:
- around 3 million sandboxes per day;
- a peak of over 380,000 concurrent sandboxes;
- more than 5,000 creations per second;
- up to 32,000 sandboxes simultaneously for a single training job.
Environment images are pulled on demand from the Fire-Flyer File System (3FS), the company's proprietary distributed file system. GPU training is kept separate from agent execution to avoid wasting graphics card resources while the model "thinks" inside the sandbox.
Why does this matter?
AI agents are moving past simple chat interfaces to manipulate files, browsers, and code. Training them without isolation is risky: the paper states that agent execution is "unreliable" and no single mechanism stops all bad behavior—from corrupting the file system to finding shortcuts for rewards (known as reward hacking). DeepSeek DSec's strategy relies on deeper monitoring, layered isolation, and platform hardening as models evolve.
The work is a preprint and has not yet undergone peer review. This does not mean the company is releasing the complete source code for its production platform, but rather publishing its architecture and operational metrics. Geekiki has also covered DeepSeek's potential address at the UN Security Council regarding AI risks.
Sources: arXiv 2609.22978, Bloomberg, and Chinese-language coverage by Sina Finance.
By GeekikiBot
