Tencent AI Executive Warns of Risk of AI Enabling Bioweapon Development

Tencent's Chief AI Scientist, Vinces Yao Shunyu, has warned of the risk that AI could enable bioweapon development in the near future. He further argues that in response to this risk, AI models should not be completely closed off, but rather an intermediate approach between openness and closure is necessary.

In the debate surrounding AI safety, the conflict persists between whether models should be made publicly available or managed by corporations. Open source models, which anyone can use and modify, have safety filters in place, but these are known to be relatively easy to remove using a technique called "abliteration." On the other hand, models operated privately, like Anthropic's 'Claude,' often impose excessive restrictions on inquiries with safety concerns, according to Yao.

Facing these challenges, Tencent is attempting to adopt its own approach. In its in-development open-weight model 'Hy4' (a format that makes model weights public), the company adopts a policy of excluding dangerous hacking knowledge from the training data stage onward, rather than adding safety filters after the fact. In essence, this is the idea of removing harmful knowledge from the "textbook" that the model learns from.

This method differs fundamentally from the current mainstream safety measures. Many models attempt to ensure safety by adding filters after training, but this runs the risk of having those filters bypassed by techniques such as abliteration mentioned earlier. By contrast, an approach that controls the training data itself is positioned as a more fundamental countermeasure in that it makes it difficult to "remove the filter later."

The combination of AI and bioweapons is a topic that has long concerned security experts and researchers. As AI becomes capable of learning and generating advanced scientific knowledge, the risk of inadvertently or deliberately providing synthesis procedures for dangerous biological materials is becoming increasingly realistic. How major AI companies address this issue is also a critical question that will shape the ethical framework of technology development.

Yao's remarks at Tencent can be seen as presenting a third perspective to the AI safety debate—rather than the false dichotomy of "openness or closure," it is controlling risks from the design stage onward. Going forward, it remains to be seen how effective this approach will be following the public release of Hy4. It is likely to attract attention both within and outside the industry as a practical case study in safe AI model design.