ByteDance’s Seed security department, after its establishment, began recruiting 10 people globally the next day, focusing on researching AI agent privilege escalation and model attack and defense.

Beating AI Express (Dongcha) – ByteDance’s Seed division has recently established a dedicated safety and alignment department, and on October 14, launched a global recruitment drive for 10 researchers, with work locations spanning Beijing, Shanghai, and San Jose, U.S. The new department is led by Yan Lin, who previously oversaw alignment for the Doubao model, and reports directly to Wu Yonghui, head of Seed. Members from ByteDance’s original Responsible AI team and some Chat team members will also transfer to the new unit. Its core focus is addressing safety challenges as AI models grow more autonomous: scenarios include agents bypassing authorization when calling tools, models being tricked by malicious prompts into violating safety limits, or failing to detect and block unknown risks in time. To tackle these issues, the team plans to refine safety training workflows across pre-training, supervised fine-tuning (SFT), and reinforcement learning (RL), conduct ongoing vulnerability scans via automated red team testing, and research areas like model behavior interpretability, anomaly tracing, and high-risk capabilities including CBRN (Chemical, Biological, Radiological, and Nuclear). ByteDance has pledged long-term investment and will provide computing power resources aligned with those used for cutting-edge model research.