DeepSeek新论文公开Agent训练！梁文锋署名

TL;DR - DeepSeek disclosed DSec, production infrastructure for creating and orchestrating isolated environments at the scale required for agent training. It reportedly provisions over 5,000 sandboxes per second and supports up to 380,000 concurrent instances while addressing image delivery, resource contention, fault recovery, and agent-driven attacks.

• DSec exposes a unified Python SDK across function-call, container, Firecracker MicroVM, and full QEMU VM backends for workloads ranging from coding benchmarks to desktop automation.
• Layered EROFS images and on-demand loading from DeepSeek’s 3FS reduce rebuild costs and data transfer; an 8,192-container deployment took 35 minutes versus more than 60 minutes for cold Docker pulls.
• Memory-sharing and reclamation techniques cut peak MicroVM memory use by 40.2% and then 21.2%, while CPU scheduling controls reduced latency inflation under 50% background load from 45.2% to 17.3%.
• AppArmor and eBPF policies constrain reward hacking and data leakage, but observed exploits—including filesystem abuse, kernel crashes, network probing, and log flooding—show that sandbox security remains an ongoing challenge.