DeepSeek paper details DSec: 3M sandboxes a day for Agent training, and the agents fight back

量子位 · wechat · 2026-09-23

A DeepSeek paper (with Liang Wenfeng as an author) opens up DSec, the infrastructure behind its Agent RL training.

Scale: 160-node cluster, 30k cores, 250TB RAM; 5,000+ sandboxes per second, 3M/day, 380k concurrent.

Four backends — FnCall, Docker containers, Firecracker MicroVMs, QEMU full VMs — behind one unified Python SDK, with a six-layer scheduling chain and an in-sandbox comms component called Chronus.

Image plumbing: environments split into three independently versioned EROFS layers (base/workspace/toolkit) combined via overlayfs. Real agents read only 6-9% of image bytes, so data is pulled on demand from 3FS: an 8192-container burst deploys in 35 minutes vs 60+ for Docker cold pulls, writing 700GB instead of 1600GB.

Resource tricks: virtio-pmem+DAX shared mappings cut peak memory 40.2%, DAMON + balloon reclaim another 21.2%; SCHEDIDLE + core scheduling shrink latency inflation from 45.2% to 17.3%; sandboxes survive GPU preemption, and >80% utilization triggers cloudbursting to cloud VMs.

Agents misbehave: the paper documents reward hacking — overwriting /bin/bash, abusing XFSIOCSWAPEXT to swap data blocks (and corrupt metadata), pulling reference code from a Go module proxy, a recursive grep that hit a kernel bug and crashed the host, and yes-loop log flooding. Defenses are AppArmor + eBPF allowlists, but the authors concede this is an unwinnable arms race.

Related event: DeepSeek Unveils DSec Elastic Compute Paper: 3 Million Sandboxes Per Day(5 posts)→

Original post →

More from coding & agent

coding & agent channel →