Tencent chief AI scientist Yao Shunyu warned that future AI capabilities could create biological safety risks at a Future Science Prize symposium in Hong Kong on November 7, 2026. The South China Morning Post reported his warning; FomoEra also reported the event date. Yao was discussing a prospective risk across AI and biology, while Hy4’s reported score comes from a separate test.
A benchmark score describes performance on that test; it is not a real-world risk estimate.
What Hy4’s reported score measures
Tencent presented the general version of Hy4 on November 4, 2026. FomoEra reported that the model scored 30.4 out of 100 on Hy-BioExperimentBench, a Tencent-designed test of biological experiment design. That number is a benchmark result, not a probability of harm or a universal safety rating.
FomoEra also reported that Tencent specialists considered Hy4’s basic biology knowledge solid but found it harder to turn that knowledge into experimental plans. The score and that assessment describe performance in the context of Tencent’s evaluation; they do not measure the likelihood of a harmful real-world outcome.
What broader AI safety evidence says
The International AI Safety Report 2026, dated February 3, says general-purpose AI can provide information relevant to biological and chemical weapons development. It also reports that several developers added safeguards to models in 2025 after pre-deployment tests could not rule out meaningful assistance to novices.
The report says it remains difficult to assess how strongly practical barriers constrain people who might misuse such information. These findings concern general-purpose AI, rather than Hy4 specifically.
The trade-off around open models
Open models may help defenders, while biological capabilities that support defense can be difficult to distinguish from those that could support offensive use. The Decoder attributed that view to Yao. Reports published November 9, 2026, also said Tencent planned to release Hy4’s general-version weights with safety-evaluation documentation.