DeepSeek reveals innovative method for training AI agents with massive sandbox infrastructure

DeepSeek published an arXiv paper detailing DeepSeek Elastic Compute (DSec), a sandbox infrastructure system that generates over 5,000 isolated training environments per second, or roughly 3 million per day per production unit. The paper, co-authored by about 130 contributors, describes four isolation backends (FnCall, Docker, Firecracker microVMs, and QEMU virtual machines) accessed through a single Python SDK, with each production unit running on approximately 160 nodes with around 30,000 CPU cores and 250 TB of memory, and peak concurrency exceeding 380,000 simultaneous environments. The research team stated that "no single mechanism can prevent all agent misbehavior" and identified failure modes including filesystem corruption and resource exploitation, addressing them through containment, observability, and ongoing safety enhancement.

Photo: U.Lucas Dubé-Cantin / Pexels DeepSeek reveals innovative method for training AI agents with massive sandbox infrastructure The Chinese AI lab's new system can spin up 3 million isolated training environments per day, tackling one of the hardest problems in building reliable AI agents. DeepSeek https://cryptobriefing.com/markets/deepseek/ just published the blueprint for what might be the most ambitious AI training infrastructure anyone has built to date. The Hangzhou-based AI company submitted a research paper to arXiv detailing a system called DeepSeek Elastic Compute, or DSec, that can generate over 5,000 sandboxes per second for training AI agents at scale. That adds up to roughly 3 million isolated environments created every single day per production unit. What DSec actually does Training AI agents is fundamentally different from training a chatbot. An agent doesn’t just generate text. It takes actions: writing code, manipulating files, browsing the web, executing commands. Every one of those actions carries risk, which means you need to contain each agent in an environment where it can’t do real damage. DSec solves this by offering four distinct isolation backends, all accessible through a single Python SDK. The lightest option, called FnCall, handles stateless operations. Docker containers provide a step up in isolation. Firecracker microVMs offer even stronger boundaries. And full QEMU virtual machines deliver the heaviest containment available. A single production unit runs on approximately 160 nodes, packing around 30,000 CPU cores and 250 TB of memory. The paper, co-authored by roughly 130 contributors, describes a custom distributed filesystem called 3FS that handles layered, on-demand image loading, keeping sandbox creation fast enough to sustain that 5,000-per-second throughput. Peak concurrency tops 380,000 isolated environments running simultaneously. AI, tech, and the markets they move—in one daily briefing. Daily. Free. Join 34,000+ readers across crypto, finance, and policy. The safety problem no one has solved Perhaps the most candid admission in the paper is a deceptively simple sentence: “no single mechanism can prevent all agent misbehavior.” The research team identified specific failure modes including filesystem corruption and resource exploitation, scenarios where agents find unintended channels to affect systems outside their sandbox. Rather than claiming to have a silver bullet, DeepSeek’s approach combines three strategies: containment the layered isolation backends , observability monitoring what agents do inside their sandboxes , and ongoing enhancement of safety measures as new failure modes emerge. Why infrastructure is becoming the real AI battleground The DSec paper signals a subtle but important shift in how the AI race is being fought. For the past few years, most of the attention has gone to model architecture and training data. DeepSeek itself made waves with its reasoning models that challenged the assumption that cutting-edge AI required cutting-edge budgets. Now the company is making the case that infrastructure for agent training deserves just as much focus as the agents themselves. The 10,000-word paper reads less like a model announcement and more like a technical manifesto for how to scale agent development responsibly. The sheer hardware requirements, 30,000 CPU cores and 250 TB of memory per production unit, suggest this isn’t something a startup can replicate in a weekend. Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy https://cryptobriefing.com/editorial-policy/ .