Tencent's Chief AI Scientist Warns AI Bio Capabilities Will Soon Make Harm Possible - and Says Anthropic Has a Strong Case for Closed Models

Yao Shunyu, Tencent's chief AI scientist, has a warning for the open-weight community: biological capabilities in frontier models are close to the point where they can actually cause harm, and developers who release weights publicly need to take dual-use evaluation far more seriously. Speaking on Saturday at a science symposium co-hosted by the Hong Kong Academy of Sciences and the Future Science Prize, Yao said that "the cyber abilities of models, and soon the bio ability of models, will reach a threshold where it's actually possible to do harm."

Biological risks, he argued, may prove harder to manage than the cyber threats the industry has been wrestling with, because the boundary between attack-relevant and defense-relevant capabilities is even blurrier in biology. His remarks were reported by the South China Morning Post.

The most striking part of Yao's talk was what he conceded to the closed-model camp. US companies such as Anthropic, he said, have made a "strong argument" that frontier models should stay closed to prevent significant societal harm. But he stopped short of endorsing a fully closed future, arguing that a mix of closed and open models may serve society best - because open weights also arm defenders. He cited a recent episode in which Hugging Face used a Chinese open model to help defend against a cyberattack after closed US systems, bound by their built-in guardrails, refused the relevant requests.

Yao, 29, joined Tencent last year after a one-year stint at OpenAI, where he contributed to products including Operator and Deep Research. His comments land at a delicate moment for his own employer: Tencent recently released Hy4, a flagship open-source model, and said its biological capabilities remain well below the company's high-risk threshold, citing internal expert reviews that found the model unable to turn foundational biological knowledge into actionable experimental protocols. Tencent plans to publish Hy4's weights together with a comprehensive system card covering its cyber and biological evaluations in the coming days, Reuters reported.

The warning is part of a broader governance push that now includes China's own tech giants. Tencent, Huawei and Alibaba contributed to a report released at the World Internet Conference in Wuzhen warning that advancing AI capabilities in tasks like pathogen design are creating major public health risks. China has proposed export controls on systems that cross critical cyberattack thresholds, and its disarmament ambassador warned at the United Nations in October that AI-biology integration lowers the barriers to attacks. US and Chinese officials are scheduled to discuss AI risks and incident notification mechanisms later this month.

Yao also said the harder problems are no longer purely technical. Solving the evaluation and safeguarding questions is "a matter of time," he argued; the bigger challenge is organizational - how labs structure research workflows, team collaboration and pre-release capability reviews. For an industry that has spent two years debating whether open weights are a gift or a weapon, the message from one of China's biggest open-weight publishers was notably even-handed: both camps are right, which is precisely why the evaluation infrastructure matters more than the license.