AI models will soon reach a threshold where real harm in biology becomes possible, Tencent chief AI scientist Yao Shunyu said at a science event in Hong Kong on Saturday. The warning, reported by the South China Morning Post, puts biology next to cybersecurity, where he said models can already cause damage. The statement matters because it came days after Tencent launched Hy4, its new flagship model, on 4 November.
What Yao said after the Hy4 launch
Yao, 29, joined Tencent last year after a year at OpenAI, according to the report. Speaking in Hong Kong, he said Chinese companies releasing open models that anyone can download and modify need to take safety very seriously. He added that biology risks may prove harder to manage than cyber risks. The remarks frame openness as both a distribution choice and a safety obligation for model providers.
Hy4 was presented with a deliberate change in training composition. Tencent said it left material on exploiting software flaws out of the training data and trained the model instead on defensive security data. The company described Hy4 as having strong basic knowledge of biology while remaining weaker than leading models at designing experiments. That assessment was based on internal expert reviews, and TNW noted it had not tested the model.
The launch sits inside a wider debate about open and closed releases. Yao said US firms such as Anthropic make a strong case for keeping top models closed. At the same time, he argued that a mix of open and closed models may work best because open models can also help defenders. Tencent plans to release Hy4 weights and a report on safety testing in the coming days, Reuters reported.
What the biology warning means for business
For companies building products on open models, the near-term consequence is tighter scrutiny of what can be downloaded, fine-tuned and redeployed. A small team using an open-weight release for support automation or research assistance may face new usage terms, access controls and documentation duties. A large enterprise with regulated data and broader deployment will likely need formal review of model provenance, modification logs and incident response for misuse.
The limits of the current disclosure deserve attention in procurement. Tencent has not yet published the Hy4 weights or the safety testing report, and its biology assessment rests on internal expert reviews rather than independent evaluation. Buyers should ask which exploit-related content was excluded, how defensive security data was selected, and how experiment-design capability was measured. This news alone does not establish that Hy4 or rival systems have crossed a biological harm threshold.
The marker to watch is the coming Hy4 release package with weights and safety test results. Hy4 already ranks as the second-best Chinese model with benchmarking firm Artificial Analysis, according to SCMP. If the report details test methods and residual risks, corporate users will gain a clearer basis for comparing open and closed options in security-sensitive work.
