ByteDance Seed Forms Dedicated Model-Alignment Preparedness Team to Tackle Agent Safety

IT Home reported on October 13 that according to a report by Sina Tech today, ByteDance Seed recently sent out an internal email announcing the establishment of the "Seed Model-Alignment Preparedness" department. It will be led by Yan Lin, who will report to Seed head Wu Yonghui, and members from the former Responsible AI team as well as parts of the Chat team will be transferred to the new department. According to the report, the establishment of this department aims to strengthen investment in safety and alignment. The core of alignment is to ensure that the model's behavior remains consistent with human intent and safety constraints, thereby enhancing the safety capabilities of Agent Models and guarding against frontier capability risks such as CBRN (chemical, biological, radiological, and nuclear). ByteDance Seed is the first frontier AI laboratory in China to establish a dedicated safety and alignment department. The safety risks of large models are becoming increasingly prominent, with incidents such as cyberattacks triggered by agent escapes occurring frequently. Based on previous reports by IT Home, on the evening of September 30 local time, OpenAI disclosed that its AI agents may have attempted to bypass security protections without authorization, potentially causing negative impacts on the systems of over a hundred organizations. The information disclosed by OpenAI shows that the scope of known out-of-control AI agent behaviors has expanded further. The company has issued notices to more than 100 third-party organizations, informing them of the discovery of "agent behavior deviating from expectations." On October 9 local time, Anthropic disclosed that its AI agents had attempted to access multiple websites of U.S. federal, state, and local governments without any human instruction. Anthropic did not release the names of the government agencies involved and has briefed the White House on these incidents. Microsoft CEO Satya Nadella published an article on October 11 stating that in an era of booming AI, simple strategies like using superintelligence to monitor itself are not enough. Nadella believes that we should take a step back and re-evaluate the trust architecture for this new era. Model providers simply cannot outsource responsibility, nor can they treat superintelligence as a set of nested black boxes, simply accepting or rejecting its recommendations, answers, and actions. We must build closed systems where its behavior can be observed, its limits tested, and its actions always contained.