Who doesn’t want AI technology to become more accessible? It can make possible for smaller research teams and independent scientists to work on problems that require enormous resources. But there is another side to this democratisation of AI or more precisely, open-weight models. The way they could help researchers discover new medicines may also enable people with malicious intent to develop dangerous things like bioweapons. Tencent‘s chief AI scientist and former OpenAI researcher Vinces Yao Shunyu believes the world may be approaching that point sooner than many expect.
Speaking at a science symposium in Hong Kong on Saturday, Yao warned that AI models are getting closer to capabilities that could cause real harm in biology and cybersecurity. “The cyber abilities of [AI] models and soon [the] bio ability of models will reach a threshold where it’s actually possible to do harm,” he said.
It is worth noting that his warning comes at a time when AI is moving beyond chatbots, coding assistants and image generators. For instance, researchers are now increasingly using the technology to understand proteins, explore biological processes and accelerate drug discovery. But if these capabilities become accessible to almost anyone, the risks could grow alongside the benefits.
Why open AI models could become a double-edged sword
Now, the debate is particularly relevant when we talk about the open-weight AI models. Unlike closed systems, whose underlying model weights remain restricted, open-weight models can be downloaded and modified by people with sufficient computing resources.
That openness has its advantages. Like, independent laboratories could use AI to explore scientific questions without depending entirely on large technology companies. Cybersecurity researchers could also customise models to identify vulnerabilities and defend their systems.
But it is also important to keep in mind that the same freedom could be exploited. People with harmful intentions might attempt to use increasingly capable models to develop biological weapons or make dangerous pathogens easier to design. This is what makes biological AI particularly difficult to manage, as the technology can potentially serve both researchers and attackers.
Yao said biological capabilities could be more complicated to regulate than cybersecurity, partly because defensive and offensive applications are harder to separate.
“All the Chinese or any open model company needs to be very responsible and very serious about those security and safety concerns,” he said.
Tencent wants AI scientists
Tencent is already exploring how AI could conduct scientific research independently. The Chinese technology company wants to build autonomous science labs where AI systems can carry out complicated experiments from beginning to end, including work involving protein folding.
Its latest flagship model, Hy4, also reflects the company’s approach to balancing accessibility with safety. Tencent says it removed the model’s hacking capabilities while retaining cybersecurity defence functions. Its biological capabilities, meanwhile, remain below the company’s high-risk threshold.
The company plans to release Hy4’s model weights alongside a system card explaining its cybersecurity and biological safety evaluations.
Yao believes neither completely open nor entirely closed AI models may offer the ideal solution. He pointed to an incident in which Hugging Face reportedly used a Chinese open model to defend against a cyberattack after closed US models refused assistance.
“Maybe the best way forward is still a mix of closed and open-source because you still want to empower everyone to leverage the technology,” he said.
– Ends
Leave a Reply