Tencent's AI chief warns of biology risks as it readies an open Hy4 model

Tencent's chief AI scientist, Yao Shunyu, has cautioned that AI models will shortly be able to do damage in biology, just as they already can in cybersecurity. He made the remarks on Saturday at a Hong Kong science event, according to a South China Morning Post report cited by The Next Web.

"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," Yao said, as quoted by The Next Web citing the SCMP.

What did Yao argue?

Yao, who is 29, moved to Tencent last year following a year at OpenAI. He said Chinese companies releasing freely downloadable open models must take safety very seriously, and that biology risks may prove harder to manage than cyber risks. He acknowledged that US companies such as Anthropic argue persuasively for keeping their leading models closed, but he argued that combining open and closed models may work best, since open ones can also help defenders, The Next Web reports.

What is Tencent doing with Hy4?

Yao spoke days after Tencent released Hy4, its new flagship model, on November 4, amid growing global concern about AI-powered hacking. Tencent said Hy4's cybersecurity training was confined to defensive work, like spotting and repairing software flaws, and kept out training material on creating hacking tools, intruding into systems or probing them for weaknesses, Reuters reported.

According to Reuters, Tencent said internal security specialists found no sign that Hy4 raised the risk of helping users break into software or get around security protections, and that its testing showed the model could not meaningfully make harmful biological experiments easier. Citing Tencent's launch announcement, The Next Web says the model has strong basic biology knowledge, though it lags leading models at designing experiments, based on internal expert reviews; The Next Web said it had not tested the model itself.

Tencent plans to release Hy4's weights later, along with a report on its safety testing, Reuters reported; The Next Web, citing Reuters, said this was expected in the coming days. "We recognise the potential safety risks posed by advanced cyber and biological capabilities, and we are committed to a more responsible approach to model training, open-sourcing, and release," the company said. Benchmarking firm Artificial Analysis ranks Hy4 as the second-best Chinese model, the SCMP said, according to The Next Web.

Why does it matter?

Chinese AI companies have in recent months shown signs of putting more into model safety despite fierce domestic competition and rivalry with the US, Reuters reported, noting that Beijing-based Moonshot AI last month limited the cyber-offensive capabilities of its newest model.

The Hy4 release also came after Tencent disclosed last month that agents it ran to evaluate other companies' models had browsed the web in unintended ways; researchers said Tencent Cloud-linked agents had queried Alibaba's Amap service. Tencent said its own Hunyuan AI models were not the source of the activity, that it saw no sign of deliberate targeting or coordination, and that it would tighten controls on agents' internet access and monitoring, according to Reuters.

The safety claims about Hy4 are Tencent's own; the promised safety report and open weights could allow outside researchers to examine them once published.