ByteDance Seed sets up a standalone Model-Alignment & Preparedness department reporting to Wu Yonghui, focused on agent safety and CBRN risk, and opens global hiring for 10 researchers

Why it matters

【Key points】

On Oct 13, The Paper, Caixin and others reported that ByteDance Seed, in an internal email, set up a standalone "Seed Model-Alignment & Preparedness" department. It is led by Yan Lin, previously Doubao's alignment lead, and reports straight to Seed head Wu Yonghui. The Responsible AI team and some Chat staff move in.

It has two priorities: keeping agents within user intent and authorization when they use tools, and guarding against CBRN risks. Its remit covers model specs, safety evaluation, alignment training, safeguards and monitoring, with more public work with third-party evaluators. Caixin says ByteDance supplied the information; outlets call Seed China's first frontier lab with a standalone safety and alignment department.

On Oct 14, Seed opened hiring for 10 researchers in Beijing, Shanghai and San Jose. The work covers safety training at every stage from pretraining to RL, automated model-vs-model red-teaming, and catching unknown risks.

【Verifiable】

The Paper: https://www.thepaper.cn/newsDetail_forward_34241236

Caixin: https://m.caixin.com/m/2026-10-13/102492839.html

iFeng (hiring): https://original.ifeng.com/c/8xDKDRxcvOS

【Our read】

1. What's new is where the team sits: safety now reports to the head of model R&D, not corporate compliance. "Preparedness" borrows OpenAI and Anthropic's frontier-risk language.

2. SemiAnalysis found only 9 of 857 models from nine top Chinese labs had a public safety evaluation (e/780). The test: does Seed's next flagship ship with a system card, CBRN evals and third-party results?

3. This is separate from the Doubao opt-out list (e/861), which was a permissions fight. Here the lab is building safety ahead of agents acting on their own.

4. Sources are media citing the email plus ByteDance's statements; no announcement on Seed's site yet, so: supporting.