← Back to Home

AI Accelerates Drug Discovery but Also Nears a Biosafety Threshold: Tencent’s Yao Shunyu Issues a Warning

At a science summit in Hong Kong, Tencent chief AI scientist Yao Shunyu warned that models’ biological capabilities could gradually cross a threshold sufficient to cause harm. But this risk assessment must still be clearly distinguished from experimentally demonstrated capabilities.

By SURL BioNews

Artificial intelligence that can help discover new drugs may also make dangerous biological knowledge easier to apply. As AI enters life sciences research, assessing its benefits and harms is becoming another challenge beyond its role as a research and development tool. Tencent chief AI scientist Yao Shunyu recently warned in Hong Kong that models’ biological capabilities may soon cross a threshold sufficient to cause real harm.

The South China Morning Post reported on November 8 that Yao made this assessment at a science summit jointly organized by the Hong Kong Academy of Sciences and the Future Science Prize. The report juxtaposed AI’s potential to accelerate drug discovery with the possibility of its malicious use to create biological weapons and trigger pandemics, highlighting the risks of open-weight models in particular. This was a warning about the direction of capability development, rather than a report of a biosafety incident that had occurred.

The Future Science Prize’s official agenda confirms that the event took place at the Hong Kong Science Museum on November 7, 2026. Yao attended as Tencent’s chief AI scientist, delivered a keynote titled “Tools, Agents, and Agents That Build Tools: The Evolution of AI for Science,” and participated in an artificial intelligence panel discussion. The agenda verifies his attendance and presentation topic but does not provide a verbatim account of this warning; the relevant statements remain based on media reporting.

For drug research and development, AI’s appeal lies in helping researchers find clues in large volumes of data and narrow the scope of subsequent experiments. However, proposing candidate directions and demonstrating that a drug is effective and safe are separate stages. A 2024 World Health Organization discussion paper also noted that AI can be applied across the development and deployment of drugs and vaccines, but that public health interests and appropriate governance must be ensured alongside commercial benefits.

Open weights are a key part of this discussion: they give users access to model parameters so they can deploy or modify models themselves, while making it harder for developers releasing those models to maintain oversight of subsequent uses. This raises governance questions about how misuse risks should be assessed before release as models become capable of providing more useful biological assistance, and what evidence should determine the degree of openness. These questions cannot be judged solely by how fluently a model responds.

Currently available reporting from the event and the official agenda do not provide specific models, training datasets, biosafety test results, or experimental validation supporting this threshold assessment, nor do they provide a timeline on which to base a risk estimate. Yao’s remarks therefore serve as an early warning but are insufficient to demonstrate that AI can already independently carry out dangerous biological procedures. What needs to be clarified next is which capabilities have been measured, how far they can go under real experimental conditions, and whether the corresponding safeguards are effective.

References

  1. News Karnataka
  2. South China Morning Post
  3. Future Science Prize Foundation