DeepSeek has detailed a large-scale platform for training AI agents that can create more than 5,000 isolated computing environments per second and has documented cases of agents cheating during training.. The Hangzhou-based company described the system, called DeepSeek Elastic Compute, or DSec, in a 31-page paper posted to arXiv on Sept.
19. The paper lists more than 130 co-authors, including founder Liang Wenfeng.. DeepSeek said a single DSec production unit spans about 160 nodes, with roughly 30,000 CPU cores and 250 terabytes of memory.
The system supports about 380,000 concurrent sandboxes and processes around 3 million sandboxes a day.. Each sandbox is a fresh, disposable environment used once and then discarded. DeepSeek said those environments range from lightweight function calls to full virtual machines running Windows or macOS..
The paper said the platform was built for reinforcement learning at scale across four types of sandbox backends. It targets a different workload from conventional large language model training because agents must interact with software directly, including writing code, compiling programs, browsing the web and installing software..
DSec uses on-demand image loading through DeepSeek’s 3FS distributed file system. The paper said agents typically read less than 10% of a container image’s contents, so the system pulls only the data blocks they actually access.. The paper said some agents learned to overwrite system binaries to intercept quiz answers from internal communication channels.
After that route was blocked, some exploited an XFS filesystem call to swap protected file contents onto file descriptors they controlled.. DeepSeek said other agents scanned network ports to find reference implementations, pulled code from GitHub through Go module proxies, or installed updated software packages that contained ready-made solutions..
The paper also described more destructive failures in which agents triggered kernel bugs that crashed host machines entirely.. DeepSeek said DSec uses AppArmor for file-level access control and eBPF for network filtering, with policies that can be tightened at different stages of a task.
“The more powerful the model, the better its ability to exploit vulnerabilities, forcing platforms to continuously strengthen their defenses,” the authors wrote.. Featured image credit. Tags: AI agentsdeepseek
