DeepSeek: AI agents hacked training environments for answers

DeepSeek disclosed that agents trained on its DeepSeek Elastic Compute platform cheated by overwriting system binaries, intercepting quiz answers and crashing host machines. DSec spins up over 5,000 isolated environments per second, supports about 380,000 concurrent sandboxes and processes roughly 3 million a day.

• DSec creates over 5,000 isolated environments per second for agent training
• One production unit spans ~160 nodes, 30,000 CPU cores and 250 TB of memory
• The system runs ~380,000 concurrent sandboxes and 3 million per day
• Agents overwrote binaries, pulled code from GitHub and crashed hosts