Artificial intelligence could soon help scientists discover new medicines far faster, but the same capability could also help bad actors develop bioweapons, Tencent chief AI scientist Vinces Yao Shunyu has warned. Speaking at a science symposium in Hong Kong on Saturday, Yao said AI models are approaching a threshold where they can cause real harm in biology and cybersecurity, and he urged caution over how openly the most powerful models are shared.
A warning from the top of Tencent's AI team
Yao, a former OpenAI researcher who now leads AI research at Tencent, said the progress is arriving sooner than many people expect. He told the symposium that AI is already being used to understand proteins, explore biological processes and speed up drug discovery, work that once needed large teams and years of effort. Smaller research labs can now tackle complex scientific problems with these tools, which he described as a genuine benefit for science. But he said the same tools lower the barrier for people with malicious intent, and the risk grows as the models get stronger.
The threshold where AI can do real harm
Yao put the danger in plain terms. He said, "The cyber abilities of [AI] models and soon [the] bio ability of models will reach a threshold where it is actually possible to do harm." His point was that AI is not only getting better at writing code and finding software flaws, it is also learning the language of biology. Once a model truly masters biology, it could help design harmful pathogens or toxins, not just helpful medicines. That dual use, he argued, is the core problem facing the field today.
The open model dilemma
A large part of Yao's warning concerned open weight models, where the trained model is released for anyone to download and run. Open access spreads the benefits of AI to small teams and independent scientists, but it also removes the guardrails that a closed system can keep in place. Yao said there was a strong case for keeping the most advanced models closed rather than open, pointing to the argument made by Anthropic for restricting access to its frontier systems. Tencent's own latest flagship model, Hy4, reflects that balancing act between accessibility and safety, and the company plans to release its weights alongside a system card explaining its cybersecurity and biological safety evaluations.
What happens next
The warning lands as governments and labs debate how to govern AI without slowing down medical research. Yao's message was not to halt AI in biology, which he said holds real promise for patients, but to weigh who gets access to the most capable systems and under what safeguards. The debate now turns to screening of DNA synthesis orders, safety testing of models before release, and rules for sharing powerful models openly. For now, his central claim stands as a marker: the same AI that can find the next drug can also help build the next bioweapon, and the gap between the two is narrowing.
