Tencent's Yao warns of biological risks as AI advances

Yao's warning was not that biology-focused AI offers only danger: frontier systems could speed discovery of new drugs and other treatments.[1] But he said biological risks may be harder to manage than cyber risks.[2] Experts warn that malicious actors could use AI to produce bioweapons or potentially start another pandemic, particularly if models are released as open-weight software.[1] For teams assessing these systems, the evidence points to a need to weigh possible scientific applications alongside risks tied to model access and capability.[1]

Yao argued for a middle ground between open and closed AI models.[3] His case for keeping some models open is that they can also help defenders.[3] He also said US companies such as Anthropic make a strong case for keeping top models closed.[2] These positions frame the choice as a trade-off rather than a simple preference for one release model: professionals evaluating deployment should consider both defensive uses and potential misuse.[3][2]

Tencent launched its Hy4 model on Nov. 4, 2026.[4] The company said it left material about exploiting software flaws out of Hy4's training data and trained the model on defensive security data instead.[2] Tencent also said Hy4 has strong basic biology knowledge but is weaker than leading models at designing experiments.[2] The company planned to release the model's weights and a report on its safety testing in the days after launch; the evidence describes a plan, not confirmation that the materials were released.[2]

The Next Web noted that open models can be downloaded and changed by anyone.[2] A reported cyber example illustrates why defenders may value access: when closed US models such as Claude refused to analyze attack commands, a platform used the open Chinese model GLM-5.2 to parse the logs.[3] Yao said Chinese companies releasing open models need to take safety very seriously.[2] Artificial Analysis ranked Hy4 as the second-best Chinese model, but that benchmark position does not change Tencent's stated limitation on experiment design.[2]