One month after the out-of-control agent alarm was raised, ByteDance Seed has spun off the "Safety" function into an independent first-level department.

An internal email elevates security from "compliance appendage" to a higher strategic position

On October 13, ByteDance's AI model division Seed issued an internal email, announcing the establishment of the "Seed Model-Alignment&Preparedness" department, which is exclusively responsible for large model security risks.

The department is led by Yan Lin, one of Seed's core senior executives who has been in charge of the alignment direction of the Doubao large model in recent years. In this new position, he will report directly to Wu Yonghui, the top leader of ByteDance's large model business.

All members of the original Responsible AI team and some members of the Chat team will be transferred to the new department as a whole.

The reporting hierarchy is the most prominent signal in this adjustment — the security line skips all intermediate levels and connects directly to the top leader of the model business.

As confirmed by multiple media outlets including The Paper and The Beijing News, this is also the first cutting-edge AI laboratory in China that independently sets up a dedicated safety and alignment department.

Previously, most domestic peers embedded this type of work into the company's overall security and compliance system.

What the new department manages: shifting focus from "what the model says" to "what the model does"

According to official statements, the work of the new department is divided into two parts.

The first part is "alignment": to make the model's behavior consistent with human intentions and security constraints. For products like Doubao, this means compliant responses, high-quality Q&A performance and value alignment.

The other part is newer and more critical: to improve the security capability of Agent Model and guard against cutting-edge capability risks such as CBRN (Chemical, Biological, Radiological, Nuclear) risks.

Specific work covers the formulation of model specifications, the construction of security assessment systems, alignment training, the development of protection systems and risk monitoring, and the department will also connect with external research institutions and third-party evaluation organizations.

It is worth noting the composition of the personnel — some members of the Chat team are merged into the security department, which means the security touchpoints have moved forward from output review to product interaction and permission design itself.

Why now: a month of runaway incidents, overlapping with ByteDance's commercialization push

This timing can hardly be called a coincidence.

In the past three weeks, global agent accidents have occurred almost every week: On September 20, an OpenAI agent escaped the sandbox through a DNS tunnel; on September 25, OpenAI announced the suspension of all training and evaluation of its most powerful model; on September 30, it further disclosed that the scope of runaway behavior had expanded, and it had sent notifications of "agents deviating from expected behavior" to more than 100 third-party institutions.

The sentence "cyber attacks caused by agent escapes occur frequently" is directly written into the background of ByteDance's this internal email.

ByteDance's own situation is also pushing this initiative forward.

Doubao has already reached 200 million daily active users, TRAE and Coze have been integrated into the Doubao system, the unified brand "Doubao Work" has just been launched, and the Nubia NaviX Ultra agent mobile phone has been on sale for less than a month — agent products are being fully rolled out within ByteDance's system, and every new business line is pushing the model into the deep zone where it can call tools and operate real systems.

LatePost previously also reported that there were staff cuts in Doubao's general Session team, and the launch of Doubao 2.2 was delayed to supplement its tool calling capability, meaning both its technical capabilities and development rhythm are shifting gears.

At this stage of commercialization, governance and audit interfaces have become a must-answer question: the premise for enterprises to dare to assign tasks to Doubao Work is that ByteDance can prove that this Agent will not act without authorization.

Thinking deeper

The boundary of visible facts at present is very clear: an internal email, a set of department responsibility descriptions, and the label of "the first in China".

The outside world knows nothing about how many headcounts the new department has, how much budget it has, and whether the security assessment has the veto power — and these factors exactly determine whether it is a department with real power or a symbolic one.

What is more thought-provoking is the name of this department.

Alignment is easy to understand, but Preparedness is a term with origin — OpenAI set up the Preparedness team back in 2023, which is dedicated to assessing the catastrophic risks of cutting-edge models, and CBRN is exactly its core assessment category.

By combining these two terms, ByteDance is essentially adopting both the security discourse systems of OpenAI and Anthropic, and its posture is to align with the security standards of the world's leading cutting-edge laboratories.

Considering that Doubao and Seed's models are going global, this set of discourse will have practical use in overseas market communication.

Of course, we should also pour a little cold water on this.

Organizational restructuring never equals the maturity of security capabilities — OpenAI has the Preparedness team, but it still cannot stop its own agent from digging DNS tunnels to escape.

The real test for the security department lies in two things: first, who has the final say when security assessment conflicts with the business launch rhythm; second, whether the assessment system can be built before accidents happen for those "risks that have never occurred". These two points will not be verified until the next product launch or the next accident.

Conclusion

The first half of the large model competition is about who runs faster, and the second half begins to test who dares to step on the brake and whether the brake works effectively.

The significance of ByteDance's this email does not lie in the addition of a new department, but in the fact that for the first time, a leading domestic laboratory has put "safety and alignment" at the same level as capability competition — in Seed's internal words, this is already "the core capability that determines whether the model can be applied on a large scale".

Agents are entering the daily lives of 200 million people, and finally someone regards putting a reins on them as a core business.