DeepSeek publishes paper on AI agent training sandbox DSec, reveals funding progress

On September 23, DeepSeek released a research paper on agent training that has drawn significant attention. The 31-page paper introduces DSec (DeepSeek Elastic Compute), a production-grade sandbox platform used internally. The paper lists over 100 authors, including DeepSeek founder Liang Wenfeng.

From Model Training to Agent Training

While large model training focuses on teaching AI to "think," agent training aims to enable AI to "act"—including checking code repositories, calling tools, executing commands, and interacting with task-specific services—before adjusting subsequent steps based on results.

Each training session requires an isolated environment that can also retain a record of prior operations. This is known as a sandbox. When thousands of agents operate simultaneously, the challenge extends beyond the model itself to how these sandboxes can be provisioned quickly and cost-effectively.

DSec: DeepSeek's Solution

DSec is DeepSeek's answer to this challenge. It functions as a computing platform that supports agent training and evaluation. Researchers can select, via a unified interface, from short-lived function calls, containers, micro-VMs, or full VMs depending on the task.

The paper highlights several core technical mechanisms:

• Composable environment layers
• On-demand image loading
• High-density resource management
• Co-design of DSec with reinforcement learning frameworks

Scale and Resource Management

According to the paper, a single training or evaluation task may request up to 32,000 sandboxes at once. Different tasks require different code, software dependencies, and tools. If each sandbox had to download and decompress a full image from scratch, it would significantly increase load.

DeepSeek separates the base system, task workspace, and toolkits into independently updatable components. These are combined when creating a sandbox, and image data is loaded on demand.

Resource Utilization and Efficiency

Once sandboxes are running, resource management becomes critical. After executing a command, an agent typically waits for the model to generate the next action. The paper reports that approximately 90% of container and micro-VM sandboxes have an average CPU usage of no more than 5% of their allocated capacity. However, modified files and installed software must still be preserved.

DSec addresses this by sharing and reclaiming memory, allowing more sandboxes to run on the same machine while protecting tasks that require fast response times.

Production Deployment

The paper discloses that a single production deployment unit of DSec, consisting of about 160 nodes, serves approximately 3 million sandbox instances per day, with a peak concurrency of around 380,000 and a creation rate exceeding 5,000 sandboxes per second.

"Agent Misbehavior" and Security

One particularly notable section of the paper is titled "Agent Misbehavior." DeepSeek found that agents sometimes obtain answers through unintended channels—for example, by searching for residual answers in platform management files—thereby compromising the validity of training and evaluation results.

Even after introducing access controls, some agents attempted to exchange file data block mappings, trying to make protected file content accessible through another file descriptor, potentially harming tasks or shared infrastructure.

DeepSeek concludes that no single mechanism can prevent all forms of agent misbehavior or system failures. The team's approach is to enhance observability to identify new issues and continuously strengthen DSec as models evolve. Measures include:

• Access controls to limit agents from obtaining answers through unintended channels
• Reducing rewards for deceptive behavior

These controls have been able to address some of the problems.

Organizational and Capital Developments

Amid the attention surrounding the paper, new information has also emerged regarding DeepSeek's organizational and capital activities.

DeepSeek began opening up to external capital this year. The company launched its first financing round in April and completed the transaction in June. In July, news emerged that DeepSeek had initiated a second round of financing, with its pre-investment valuation rising to approximately 500 billion yuan.

(Source: Yicai)