Clarity for AI buildersSponsored by AI Routing
adoption

Tencent’s AI chief warns of bioweapon risk and calls for a mix of open and closed models

Tencent’s AI chief warns of bioweapon risk and calls for a mix of open and closed models
Tencent’s chief AI scientist warns that AI could soon help develop bioweapons and says neither open nor closed models suffice. Vinces Yao Shunyu, formerly of OpenAI, argued for a mix of open and closed models so defenders can use the technology too, The Decoder reports, citing the South China Morning Post. Yao cautioned that offensive and defensive uses are harder to separate in biology than in cybersecurity. He also urged caution across the industry, saying the capability gap between Chinese and US models is small. Yao pointed to a breach at Hugging Face to illustrate the value of open models. According to The Decoder’s account of the incident, Anthropic’s Claude refused to analyze attack commands, while the open model GLM-5.2 helped investigators interpret the logs. The model did not stop the attack: Hugging Face’s security team patched the vulnerability, rebuilt affected servers and rotated credentials. The case shows how an open model can help with a specific defensive task that a hosted model declines. It does not show that open models generally favor defenders, or that hosted models should drop their safety filters. Hugging Face itself cautioned against the latter conclusion. Anthropic CEO Dario Amodei disputes the broader claim that open models benefit defenders more than attackers. As The Decoder reports, his concern is that safeguards are difficult to enforce once model weights are public and cannot be recalled. Anthropic has not called for an outright ban on open models. The Decoder cites a technique called abliteration, which can remove trained refusal behaviors from models whose weights are available. Research cited in the article indicates that existing countermeasures can make this harder but do not prevent it. Hosted models give providers more control over what they serve, but that control can also block legitimate work. The Decoder reports that Claude routed some flagged cybersecurity and biology requests to the less capable Opus 4.8. Artificial Analysis estimated that the routing initially affected 8 to 9 percent of tasks; that figure measures how often requests were rerouted, not how many were wrongly flagged. Users reported that even MRI analysis was flagged, and Anthropic acknowledged that its filters were too aggressive. Tencent’s planned open-weight Hy4 model illustrates a different approach. The Decoder reports that Tencent plans to release the model with a safety report. Tencent says Hy4 falls below its high-risk threshold for biological capabilities and that it filtered the training data to remove hacking capabilities. The distinction is that training-data filtering aims to keep a capability out of the model, whereas a refusal filter can leave the capability in place and block its use. The Decoder cites researchers who say training-data filtering can withstand abliteration, but the supplied reporting does not independently verify Tencent’s biological-risk assessment or establish how much cyber capability the filtering removed.