Breaking

DeepSeek Details DSec Sandbox Platform for Large-Scale AI Agent Training

September 23, 2026 at 15:55 EDT

AI Agents
Infra & Chips
Research & Papers

DeepSeek has published a detailed account of DeepSeek Elastic Compute, or DSec, a production sandbox platform designed to run the isolated environments needed to train and evaluate AI agents at scale. The disclosure shifts the focus from model weights to the machinery that lets agents practice tool use, coding, security and computer-use tasks while limiting some of the failure modes that emerge when models act inside real systems.

Sep. 19, 2026 - DeepSeek

A sandbox factory for AI agents

DeepSeek's DSec platform is built to run millions of isolated, stateful practice environments while containing some forms of agent misbehavior.

sandboxes served per day

380K+

concurrent sandboxes in production

5,000+

sandbox creations per second

160

nodes in one production unit

Environment setup got shorter

Shared EROFS layers replaced repeated tar extraction in one provisioning test.

79 min

tar extraction

45 min

EROFS layers

Takeaway: The reported run was 1.76x faster, with far less disk-write pressure.

Agentic RL loop

1. Rollout

Agents act in stateful sandboxes.

2. Reward

Tests, exit codes and verifiers score work.

3. Update

Policy training uses collected trajectories.

Why builders care

The disclosure gives concrete production numbers for agent training, plus real failure cases around answer leakage, resource exhaustion and filesystem damage.

What remains open

The report is not peer reviewed, backend selection is still the user's responsibility, and the access controls are only partial defenses against agent misbehavior.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.