Skip to main content
ModelsResearchSecurity

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

Source: The Decoder (opens in a new tab) · Maximilian Schreiner

Intel Summary

Tencent Chief AI Scientist Vinces Yao Shunyu stated that artificial intelligence will soon possess the capability to assist in developing biological weapons, arguing for an architectural middle ground between open and closed models. Yao highlighted that safety filters on open-source models can be removed through abliteration, whereas closed models such as Anthropic's Claude often over-censor. To mitigate these risks, Tencent is directly removing dangerous hacking knowledge from the pre-training dataset for its upcoming open-weight model, Hy4.

Why It Matters

Post-training alignment remains vulnerable to weight modification techniques like abliteration in open-weight models. Tencent's approach underscores an industry shift toward pre-training data sanitation to prevent high-risk capability acquisition, influencing how open-weight developers balance safety controls against model utility for cyber and biosecurity domains.

Organizations & Entities

Topics