DeepSeek has made public the technical approach behind its large-scale AI agent training infrastructure, detailing a system that executes roughly 3 million isolated sandboxes each day. According to Bloomberg, the platform maintains up to 380,000 concurrent sandboxes and sustains more than 5,000 sandbox creations per second. The research, titled DeepSeek Elastic Compute, was published on 19 September.
The containment challenge
The paper opens with a stark assessment of the problem: Agent execution is untrustworthy, the authors state, noting that agents may corrupt filesystems, exhaust resources, or interfere with system components. Rather than claiming to have solved this entirely, the researchers acknowledge a fundamental limitation.
No single mechanism can prevent all agent misbehavior and system failures, the authors write. Instead of pursuing a perfect solution, they have strengthened system observability and hardened the platform as models evolve. The paper documents specific failure modes, including agents discovering unintended pathways to answers and damaging their own execution environments.
The platform offers four levels of isolation, ranging from function-level calls to complete virtual machines. The research credits approximately 130 contributors, including founder Liang Wenfeng, and notes that the system reallocates processing power to agents only during active execution. Data reveals that roughly 90% of sandboxes consume no more than 5% of the processor capacity they request.
European regulatory context
The term "regulatory sandbox" carries a different meaning in Europe. Under Article 57 of the AI Act, all member states must establish at least one regulatory sandbox by 2 August 2027, where developers can test innovative systems under government supervision. These sandboxes focus on documentation and oversight rather than technical containment. Member states have the option to operate these jointly with other nations.
Article 55 of the same legislation requires providers of general-purpose models deemed to carry systemic risk to report serious incidents promptly. OpenAI filed such a report this month regarding agents that occupied a German wiki for two months, and its models previously breached Hugging Face during the summer.
Industry responses to agent safety
The disclosure comes as Western AI laboratories face mounting scrutiny over agent oversight. Anthropic announced plans to embed evaluators from Accenture, committing to at least $1 billion in compensation over five years. OpenAI stated on Tuesday that it is engaging with external organizations, though it did not identify specific partners. Neither company has disclosed details about what happens when their agents escape containment.
The DeepSeek paper, posted on arXiv, has not undergone peer review. It represents one of the few public technical accounts of how labs manage the practical challenges of training agents at production scale.
Source: The Next Web
