Tencent ships Hy4 AI with guardrails on hacking and bio misuse

Tencent has rolled out Hy4, a new AI model built with guardrails aimed at keeping it out of the wrong hands. The Shenzhen company says its cybersecurity training stayed strictly defensive — spotting software vulnerabilities and helping patch them. Left on the cutting room floor: anything about building hacking tools, breaking into systems, or probing them for weaknesses.

Tencent says its internal security specialists put Hy4 through its paces and found no evidence it raised the risk of users cracking software or slipping past security protections. On the biology side, the company's tests concluded the model couldn't meaningfully make harmful experiments easier to carry out. Model weights are coming later, along with a report laying out those safety tests, according to firstpost.com.

The launch lands in a jittery stretch for AI watchers. Tencent disclosed last month that AI agents it used to test third-party models had shown unintended web-browsing behavior — and researchers found agents tied to Tencent Cloud had gone hunting for information from Alibaba's AMP service, firstpost.com reported.

Tencent pushed back on that one. The activity wasn't connected to its Hunyuan models, the company said, and an internal review turned up no signs of deliberate targeting or coordinated action. It still pledged to tighten the reins on what its AI agents can reach online and to beef up monitoring. Tencent is the company behind WeChat, China's dominant messaging app with over 1 billion monthly active users.

It isn't flying solo on safety. Beijing-based Moonshot AI restricted the cyber-offensive capabilities of its latest Kimi model last month, a signal that Chinese AI firms are putting more weight on model safety even as domestic competition and the rivalry with the US both crank up. Tencent said it recognizes the safety risks that come with advanced cyber and biological capabilities, and that it's committing to a more responsible approach to training, open-sourcing and releasing models.