ByteDance Seed Creates Standalone Model Safety and Alignment Department

The Responsible AI team and some members of the Chat team will move into the new department, Leiphone reported. The move is intended to increase investment in safety and alignment, with alignment defined as keeping model behavior consistent with human intent and safety constraints.

The department will focus on improving the safety capabilities of agent models and preventing risks from frontier capabilities such as CBRN—chemical, biological, radiological and nuclear—according to the report. Safety risks around large models have become more visible, and incidents such as cyberattacks triggered by agent escapes have occurred frequently, the report said.

Previously, safety and alignment work by domestic large-model research teams was mostly folded into companies' overall security and compliance systems, Leiphone reported. ByteDance Seed is described as the first frontier AI lab in China to establish a separate safety and alignment department.

Seed's internal assessment is that as models become more capable and application scenarios more complex, safety and alignment are no longer only risk-prevention measures. They are becoming a core capability that determines whether models can be deployed at scale, according to the report.

The new department will continue to invest in several areas: model specification development, safety evaluation systems, alignment training, protection systems and risk monitoring. It will also increase public cooperation with external research institutions and third-party evaluators, Leiphone said.

The stated goal is for frontier models to complete tasks effectively within safety and authorization constraints, while reducing risks of abuse and behavioral deviation as model capabilities and autonomy increase. The department is also expected to build a comprehensive safety system for unknown risks, according to the report.