Chinese AI giants start stripping hacking skills from open models to curb risks

Background

With artificial intelligence models becoming powerful enough to autonomously develop sophisticated cyber exploits, companies face mounting risks of misuse. Unlike closed-source models where developers can monitor user activity and ban violators, open-weight models allow anyone with sufficient hardware to download and modify the software, making safety by design a critical priority.

Details

To address this, Chinese technology giant Tencent stripped its latest Hy4 model of cyber-hacking capabilities while retaining defense features, following a similar move by Moonshot AI in October. Instead of abandoning open weights, companies are selectively removing offensive cyber data during training. However, experts note a clear trade-off: Moonshot AI acknowledged that its filtered model scored 81 percent on an internal security benchmark, compared with 90.5 percent for the unrestricted version. Furthermore, cybersecurity skills are dual-use, meaning defense tools can still potentially be exploited by malicious actors.

Impact

This emerging safety practice highlights the difficult balancing act Chinese AI firms face as they try to maintain their market edge without violating tightening government export controls and risk thresholds. While targeted restrictions may weaken specialized security performance, broader competitiveness will still rely on coding quality, cost, and deployment flexibility.