Tencent strips cyber-hacking capabilities from open-weight Hy4 model, following Moonshot AI's filtered Kimi K3.1 Pro
Published · curated by AI Is Going Just Great
Source: scmp.com ↗
A capability the model never learned is usually harder to bring back than one it was just trained to refuse. But technically harder isn’t impossible.
Tencent said this month that it had stripped its Hy4 model of cyber-hacking capabilities while keeping its cyber-defence capabilities intact, following Moonshot AI's October release of a filtered Kimi K3.1 Pro. Both are open-weight models, so anyone with sufficient hardware can download and modify them and neither company can monitor usage the way OpenAI and Anthropic do with closed models. The fix is upstream: delete the offensive cyber data from the training set so the capability is never learned in the first place. The shift follows Anthropic's September report that Z.ai's GLM-5.3 could autonomously develop sophisticated exploits, and Tencent's own statement last month that AI agents on its cloud platform had displayed "unintended web-browsing behaviour" after reports they tried to extract data from Alibaba's Amap service.
Moonshot AI put a number on the cost: its filtered model scored 81 per cent on an internal cybersecurity benchmark against 90.5 per cent for a version trained without the restrictions. Analysts describe the practice as emerging rather than standard, and note that cyber capabilities are dual-use, so vulnerability detection survives the filter along with the offence it enables. Both companies are also releasing in stages, with the open-weight version arriving weeks after the closed one; Tencent has yet to ship open-weight Hy4 at all.