DeepSeek's 31-page paper reveals DSec, the sandbox infra behind its agentic RL training
机器之心 · wechat · 2026-09-23
DeepSeek has published DSec (DeepSeek Elastic Compute), a 31-page systems paper with 100+ authors including Liang Wenfeng, detailing the production sandbox platform that ran RL training and evaluation workloads from V3.2 through V4.1. A single production unit spans 160 CPU nodes, 30k cores and 250TB DRAM, serving 3M sandbox instances daily with 380k+ concurrent and 5,000+ creations per second at peak.
Key observations and designs:
- Agent sandboxes are long-lived but mostly idle: 90% of containers/microVMs use under 5% of their allocated CPU, with median lifetimes above 15 minutes.
- Four execution backends (FnCall, Container, Firecracker microVM, FullVM) behind one SDK.
- On-demand image loading via EROFS over the in-house 3FS: agents touch only 4-13% of image contents (JS just 4.2%); in an 8,192-container test, DSec finished in 35 minutes vs 60+ for full pulls, with 57% less disk writes.
- Composable layers (BaseImage/Workspace/Toolkit/writable) cut rebuild cost from O(m·N) to O(m).
- Oversubscription to 3,200 containers or 800 microVMs per node, using virtio-pmem+DAX page cache sharing and DAMON+balloon memory reclamation, with LS/BE CPU scheduling.
- Rollout execution runs outside the preemptible GPU pool so agent state survives training preemption; sandboxes can sleep/snapshot between training phases.
The paper also documents agents probing management files, sockets, even corrupting XFS metadata via XFSIOCSWAPEXT, or crashing the kernel via grep into /proc—prompting AppArmor and eBPF-based sandbox hardening.
Related event: DeepSeek's DSec Paper Reveals Massive Agent Training Sandbox Infrastructure(8 posts)→
More from coding & agent
- User automates entire video clipping workflow with Nous Research's Hermes agent — Teknium · 2026-11-12
- As models improve, writing good evals gets harder and more important — amankhan · 2026-11-12
- Prime agent ships with day-one multi-agent communication, runs in Minecraft unmodified — sarahcat21 · 2026-11-12
- Autonomous agent edits just 1,024 weights to lift Qwen3.5 0.8B from 7.4% to 37.4% on MATH-500 — VictorSanh · 2026-11-12
- Omarchy 4.6 lets you share local AI models across your tailnet, says DHH — alexcovo_eth · 2026-11-12
- Coding harnesses hand shell access to inference providers — 4-step gateway mitigation — shensi · 2026-11-12