Tencent's chief AI scientist warns AI bioweapons are coming and backs a mix of open and closed models

Tencent's chief AI scientist, Yao Shunyu, a former OpenAI researcher, says AI will soon be able to help build biological weapons, and that no model strategy, fully open or fully closed, can handle that by itself, The Decoder reports, based on comments to the South China Morning Post.

His argument

Yao wants a mix of open and closed models so defenders can use the technology too. Biology is harder than cybersecurity, he says, because attack and defense are tougher to pull apart. And since Chinese and US models are close in capability, he wants the whole industry to move carefully.

His example: during a hacking incident at Hugging Face, closed US models including Claude refused to analyze the attackers' commands because of safety filters, so the team turned to GLM-5.2, an open Chinese model, to read the logs. Hugging Face's own report supports that, with limits. The Chinese model only helped with forensics; the attack was stopped by patching the hole, rebuilding servers and rotating credentials. Hugging Face also said the incident isn't a case against filters on hosted models.

The other side

In July, Anthropic stood alone among the big labs in refusing to sign a letter, backed by Nvidia, warning against early limits on open models. CEO Dario Amodei argues open weights may be riskier, since safeguards are hard to enforce and released weights can't be recalled, though Anthropic doesn't call for a ban.

The fragility is real. A technique called abliteration strips refusal behavior out of open weights, and thousands of such modified models sit on Hugging Face. Research suggests countermeasures make it harder but don't stop it. Keeping dangerous knowledge out of training data in the first place holds up better.

Closed models have the opposite problem. Since Claude Fable 5, suspicious cyber and biology requests get routed to the weaker Opus 4.8. Artificial Analysis says that first hit 8 to 9% of tasks, users reported MRI analysis flagged as bioterrorism, and Anthropic admitted the filters were too aggressive.

What Tencent will ship

Tencent plans to release Hy4 as an open-weight model with a safety report in the coming days. It says Hy4's biology skills are under the high-risk line and that it removed hacking abilities by filtering training data, an approach researchers say survives abliteration. China's TC260 standards body is drafting a security standard for open-weight models; neither China nor the US has a binding limit on which models can be published.

Scoreboard update: Draw. AI is getting dangerous in biology, and nobody has a filter that works both ways yet.

Stay undead.