Tencent’s AI chief says AI bioweapons are coming, but closing off models isn’t the answer

What happened

Tencent’s chief AI scientist Vinces Yao Shunyu warned that AI capabilities will soon enable the development of bioweapons. He argues that shutting down or closing AI models completely is not the answer. Instead, Tencent plans a middle ground approach for its new open-weight model Hy4 by removing dangerous hacking knowledge directly at the training data stage. Yao also points out that open-source models with safety filters can be too easily bypassed, while closed systems like Anthropic’s Claude tend to block access to valuable information too aggressively.

Why it matters

AI’s ability to design bioweapons raises severe safety and security concerns that affect regulators, national security agencies, AI developers, and businesses building AI applications. Simply closing off AI models slows innovation and fragments the ecosystem but ignoring risks invites misuse. Yao’s critique of safety filters shows that superficial guardrails can create a false sense of security if they are easy to remove or circumvent. His approach to proactively excluding harmful data during training highlights a shift toward more vigilant risk management embedded in AI development rather than just reactive content controls.

For AI operators, this signals a greater need to scrutinize the provenance and filtering of training data, not just tune safety layers on top of delivered models. For investors and product builders, it means expecting AI providers to balance openness with real-world harm reduction, or risk losing trust and regulatory backlash.

What to watch next

Tencent’s Hy4 will be a test case for training-data-level filtering of dangerous content paired with open model weights. Observing its effectiveness in blocking sensitive knowledge without killing usability will be critical. The industry should watch if other major AI players adopt similar middle ground approaches instead of doubling down on either fully open or fully closed models.

Regulators may push for greater transparency on what data is removed at training time and stronger standards for evaluating the security risks of released models. Stakeholders should expect ongoing debates about the best methods to balance innovation, transparency, and public safety in AI development around dual-use threats such as bioweapons.

AI Quick Briefs Editorial Desk