Tencent and Moonshot AI are limiting cyber-hacking capabilities in their open-weight models, the South China Morning Post reported on 15 November. The change matters because users can download and modify open-weight models, putting some uses beyond their developers’ oversight. Both companies limited their models’ hacking capabilities before releasing them.
Key points
- Tencent says Hy4 retains cyber-defence capabilities after offensive material was filtered from its training data.
- Moonshot AI made a similar change to Kimi K3.1 Pro in October.
- Researchers say training-data filtering may make misuse harder, though defensive cyber skills can also serve attackers.
Tencent filters Hy4’s cyber training data
Tencent says it has removed Hy4’s cyber-hacking capabilities while keeping the model useful for cyber defence. For developers releasing model weights, that puts more of the safeguard into training rather than relying on a model to refuse harmful requests after release.
Tencent released Hy4 on 4 November after limiting its security training material to defensive content, Technology.org reported. The company said it excluded material covering exploit development, reverse engineering and penetration testing. Those techniques can also help security teams find weaknesses in their own systems, making the distinction between offensive and defensive training consequential.
The Hy4 line follows a preview that Tencent released in August. That preview used a mixture-of-experts architecture with 770 billion total parameters, of which 49 billion activated for each token. Tencent said its approach to training and releasing models reflected concerns about advanced cyber and biological capabilities.
Moonshot AI also trims Kimi K3.1 Pro
Moonshot AI made a similar change to Kimi K3.1 Pro in October. AI Affairs previously covered the model’s release and its fourth-place position on an AI index. The cyber restriction is the new element: Moonshot is also trying to narrow what an open-weight model can help users do.
The companies’ moves follow reported misuse and cyber-capability incidents. In September, Anthropic reported that Z.ai’s GLM-5.3 could develop sophisticated cybersecurity exploits autonomously. Tencent said in October that AI agents on its cloud platform had shown unintended web-browsing behaviour, after reports that the agents tried to extract data from Alibaba’s Amap navigation service.
Researchers weigh the limits of filtering
Gabriel Wagner, a researcher at Beijing-based AI safety organisation Concordia AI, said removing offensive material during training could leave a model less useful to attackers even if they bypass its built-in refusals. Luke Li Xiaojian, chief executive of Beijing-based AI safety company Fangcun AI, cautioned that capabilities useful for defence, including finding vulnerabilities, can also be used for attacks.
Wei Sun, an AI analyst at Counterpoint Research, described the Tencent and Moonshot moves as an “emerging practice” rather than an industry standard. Wider adoption, comparable tests and clear thresholds for release would be needed for a standard, Sun told the South China Morning Post.
