Tencent AI Chief Warns Frontier Models Are Nearing Biological Hazard Thresholds

On one side of the artificial intelligence industry, developers argue that open-weight models must be freely released to democratize scientific discovery and build defensive security tools. On the other, safety researchers warn that these exact systems are rapidly approaching a capability threshold where they could be used to engineer pathogens or design biological weapons.[1][2]

Yao Shunyu, Tencent’s chief AI scientist and a former OpenAI researcher, stepped directly into this tension during a rare public appearance in Hong Kong on Saturday. Speaking at a science symposium co-hosted by the Hong Kong Academy of Sciences, Yao warned that frontier models will soon master biology.[1][2]

That mastery presents both immense benefits for drug discovery and severe global security risks, Yao told the audience. He argued that the biological capabilities of current models are nearing a threshold where tangible harm becomes possible, particularly if those models are released as open-weight software without strict regulation.[1]

Yao compared the impending shift in biological capabilities to the recent alarm over cybersecurity that followed the industry's sudden leap in automated coding. However, he noted that biological risks may prove significantly harder to manage or disentangle than cyber risks, because biological knowledge is inherently dual-use.[1][2]

"The cyber abilities of models and soon the bio ability of models will reach a threshold where it's actually possible to do harm," Yao said at the symposium. "All the Chinese or any open model company needs to be very responsible and very serious about those security and safety concerns."[1]

The Hy4 Release

The warning arrived just days after Tencent released Hy4, its newest flagship large language model. The Shenzhen-based technology giant launched the 770-billion-parameter mixture-of-experts model on November 4, positioning it as a direct challenger to the world's leading open-source systems.[3][5]

Tencent markets Hy4 as a massive leap forward, heavily promoting its new visual understanding and computer-control capabilities. The company claims the model can process complex charts, conduct visual mathematical reasoning, and operate graphical user interfaces, moving beyond the text-only constraints of its August preview release.[3][5]

AI

However, independent benchmarking paints a more grounded picture of what has actually shipped. While Hy4 ranks as the second-best performing model produced in China, the benchmarking firm Artificial Analysis places it below the global average in intelligence for comparable models, noting that it is highly verbose and slower than average per task.[1][4]

The gap between marketing claims and verified capability is particularly relevant to the safety debate. Tencent acknowledges that Hy4 still trails the world's leading closed-source systems on challenging internal benchmarks for research-level mathematics, long-horizon coding, and model self-improvement.[3]

To address the exact risks Yao highlighted, Tencent applied strict filters to Hy4's training data. The company deliberately excluded material related to offensive cybersecurity exploitation, such as reverse engineering and penetration testing, to avoid amplifying the model's ability to conduct automated cyberattacks.[1][2]

Testing for Biological Hazards

Instead, Tencent restricted the model's cybersecurity training entirely to defensive applications, such as vulnerability remediation. The company claims this approach preserves the software's usefulness for legitimate security research without handing malicious actors a tool to bypass digital defenses.[2][3]

The company applied a similar evaluation framework to the model's biological capabilities. According to Tencent's launch announcement, internal security assessments found that while Hy4 possesses strong foundational knowledge of biology, it remains weak at designing complex, open-ended experiments.[1][2]

Because the model currently lacks the ability to turn foundational biological knowledge into actionable, real-world protocols, Tencent concluded that its biological capabilities remain well below the high-risk threshold. The company stated that the model could not meaningfully make harmful biological experiments easier to carry out.[1][2]

Those safety assurances currently rely entirely on Tencent's internal testing. The company has promised to release the underlying model weights and a comprehensive system card detailing its cyber and biological safety evaluations in the coming days, which will allow independent researchers to verify the claims.[1][2]

Yao argued that despite the risks, a mix of open and closed models remains the best path forward for the industry. While he acknowledged that United States firms like Anthropic make a strong case for keeping top-tier models closed, he maintained that open models are essential because they actively help defenders build better security tools.[2]

Autonomous Science Labs

The tension between capability and safety will only sharpen as Tencent pursues its broader research ambitions. Addressing the computer science students in Hong Kong, Yao outlined the company's plans to operate autonomous science labs driven entirely by artificial intelligence.[1]

These proposed laboratories would use AI systems to run complex biological experiments end-to-end without human intervention. Yao specifically cited inverse protein folding—a process critical to both advanced drug discovery and potential pathogen engineering—as a target for this autonomous infrastructure.[1]

Yao expressed confidence that the engineering hurdles standing in the way of autonomous laboratories will inevitably be cleared. "I think all the technical changes can be solved, where it's just a matter of time," he told the symposium audience.[1]

The true obstacle, Yao warned, will not be the software itself, but rather how human institutions manage it. "I think the bigger challenge is societal and organisational," he said, pointing to the difficulty of adapting regulatory frameworks to keep pace with automated scientific discovery.[1]

The push for autonomous research aligns with a broader trend among Chinese technology giants. Last week, Tencent, Huawei, and Alibaba contributed to a report on frontier AI safety at the World Internet Conference in Wuzhen, which explicitly warned that AI systems' increasing capabilities in pathogen design were creating major public health risks.[1]

Closing the Capability Gap

The urgency of these warnings is driven by the rapid convergence of global AI capabilities. Yao urged caution across the wider industry, explicitly noting that the performance gap between Chinese open-weight models and their closed-source United States counterparts is no longer large.[1][2]

That narrowing gap means that safety protocols developed in isolation by a single company or country are increasingly insufficient. If an open-weight model released in Shenzhen possesses the same biological reasoning capabilities as a closed model developed in San Francisco, the proliferation risk becomes global immediately.[1]

The dual-use nature of biological AI complicates any regulatory response. The same algorithmic reasoning required to design a targeted cancer therapeutic is often identical to the reasoning required to optimize a viral payload, making it nearly impossible to filter training data without crippling the model's medical utility.[1]

Tencent's decision to filter offensive cyber data from Hy4 demonstrates that capability pruning is possible in software engineering, where attack and defense rely on distinct code patterns. Biology offers no such clean division, leaving developers to rely on behavioral guardrails that are notoriously easy for dedicated users to bypass.[1][3]

For now, the industry remains in a precarious waiting period. The models currently available to the public possess encyclopedic knowledge of biology but lack the agentic planning required to execute physical experiments, creating a temporary buffer against catastrophic misuse.[1][2]

As companies like Tencent actively work to solve those exact planning limitations through projects like autonomous science labs, that buffer is shrinking. The release of Hy4's weights in the coming days will provide the first independent look at exactly how close the open-source frontier has moved toward the biological threshold.[1][3]

The upcoming safety report will also test the viability of Tencent's internal evaluation framework. If independent researchers find that Hy4 can be prompted to generate actionable biological protocols despite the company's safeguards, it will signal that current filtering techniques are insufficient for frontier models.[1][2]

AI

The debate over open-weight biological capabilities cannot be settled by a single model release. The threshold Yao described is not a static line, but a moving target defined by the continuous integration of artificial intelligence with physical laboratory equipment and automated synthesis tools.[1]

The immediate test of Tencent's safety claims will arrive when the Hy4 weights are published to the open-source community. Until independent security researchers can probe the model's biological reasoning directly, the industry must take the company's internal assurances at face value.[2][3]