Tencent's Chief AI Scientist Warns Models Are Approaching a Biological Harm Threshold

Tencent's chief AI scientist warned Sunday that AI models are approaching a point where they will have the technical capability to cause biological harm — and that this risk may be harder to contain than the cybersecurity threats the industry has spent years fighting. The warning came from Yao Shunyu, one of China's most senior AI researchers, speaking at a Hong Kong science event days after Tencent launched its Hy4 flagship model.

Why this matters: When a top scientist at one of China's largest AI companies publicly aligns with Western biosafety researchers on the same risk, it removes a key argument that biosecurity concerns about AI are a Western regulatory tactic rather than a genuine technical problem.

Background

The concern that advanced AI could lower the barrier to creating biological weapons or engineering dangerous pathogens has circulated in biosecurity and AI safety research circles for several years. Studies from groups including the Nucleic Acid Observatory and research cited by the U.S. Biosecurity Board have found that frontier language models can provide meaningful "uplift" — that is, genuine technical assistance — to bad actors attempting to work with dangerous biological agents. Until now, those warnings have come almost exclusively from Western academic and policy institutions.

Yao's remarks change the geography of that concern.

What Yao Said

According to the South China Morning Post, Yao said:

"The bio ability of models will reach a threshold where it's actually possible to do harm."

He framed this as a near-term development, not a distant hypothetical. He also argued that the biological risk category is structurally harder to manage than cybersecurity threats, because:

• Cyber attacks are reversible — systems can be patched, data can be restored.
• Biological harm is not — a released pathogen or a synthesized compound cannot be recalled.

Yao did not name a specific timeline or identify any existing model as having crossed the threshold.

The Technical Picture

Current frontier models — including those from OpenAI, Anthropic, Google DeepMind, and Tencent's own Hy4 — are routinely tested against biosafety benchmarks before release. AI labs have implemented filtering systems designed to refuse detailed synthesis instructions for dangerous compounds. But researchers have documented that these filters can be bypassed with prompt engineering (a technique for crafting inputs that circumvent safety restrictions), and that the specificity of information available through models has increased sharply with each generation.

The concern is not that an AI will spontaneously develop a bioweapon. It is that a person with limited expertise could use a capable AI as a knowledgeable collaborator to significantly accelerate work that would otherwise require years of specialized training.

Why a Chinese Scientist's Warning Carries Weight

AI safety discourse has been shaped largely by U.S. and European researchers, government bodies, and AI companies. Critics — particularly from China and parts of the Global South — have occasionally argued that biosecurity and existential risk framing serves to justify restricting Chinese AI development. Yao's remarks undercut that argument. A chief scientist at Tencent does not benefit professionally from amplifying Western regulatory concerns; his comments suggest the technical worry is real and crosses institutional lines.

What to Watch

Watch for whether Tencent or other Chinese AI labs begin publishing biosafety evaluation frameworks of their own, as OpenAI, Anthropic, and Google DeepMind have done. If Chinese labs voluntarily adopt similar red-teaming (adversarial testing of model capabilities) processes for biological uplift, it would signal an emerging global consensus on this risk class — and potentially create a foundation for international coordination that does not currently exist.

Hector Herrera covers AI science and safety for NexChron. Sources: South China Morning Post.