Breaking

DeepSeek Details DSec Platform for Running 380,000 Concurrent AI-Agent Sandboxes

September 23, 2026 at 13:24 EDT

AI Agents
Infra & Chips

DeepSeek has published a detailed report on DeepSeek Elastic Compute (DSec), a production sandbox platform built to run large-scale AI-agent training and evaluation workloads. The system is described as reaching more than 380,000 concurrent sandboxes in production while adding controls for reward hacking, environment tampering and other agent failures.

Sep 19, 2026 - DeepSeek

A Sandbox Factory for Agent Training

DSec is designed to keep stateful AI agents running at production scale while limiting reward hacking, unsafe access and environment failures.

380K+

concurrent sandboxes

sandboxes per day

5,000+/s

creation rate

~160

CPU nodes per unit

On-demand loading cuts startup time

8,192-container burst on a 10-node cluster

On-demand EROFS finished near the local baseline while cold eager pulling took over an hour.

The training loop DSec supports

Build

Create isolated task environments.

→

Roll out

Agents run tools and modify state.

→

Score

Rewards come from execution signals.

→

Update

Training reconnects to preserved rollout state.

Why builders care

High-density sandboxes make coding, security and browser-use agents easier to train and evaluate at scale.

What remains open

The headline scale is author-reported, and the tests measure infrastructure mechanisms rather than end-to-end model gains.

Containment is part of the product

DSec combines AppArmor, eBPF network allowlists, IAM, quotas and monitoring, while acknowledging that no single mechanism blocks every destructive agent behavior.

Continue reading

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