Moonshot curbs hacking ability in new Kimi model

Beijing-based AI startup Moonshot AI has reduced the hacking potential of its latest model, the Kimi K3.1 Pro, amid growing concerns that advanced AI could make cyberattacks easier. Introducing the new 3-trillion-parameter model, Moonshot removed cyber-related data from its training and removed potential cyber capabilities during development, while keeping the model's general coding skills.

The company plans to release the model's weights, the files allowing outside developers to run and modify it, by November 4. Moonshot aims to avoid releasing weights that would push the boundaries of cyber capability too far. As AI becomes more autonomous, there is growing worry that it could make hacking simpler and cheaper. China has proposed export controls that would ban AI-driven malware-generation technology and restrict AI-assisted tools for discovering vulnerabilities.

Moonshot claims that open-weight models can improve security by enabling more organizations to inspect and secure their own software, including older systems. However, open-weight releases make it harder for developers to control access and update safeguards once the files are widely distributed. Moonshot's K3.1 Pro scored 81.0% on its internal cyber benchmark, compared to 90.5% for a similar version without cyber-reduction measures.

The test involved finding and exploiting software vulnerabilities in a controlled environment. Despite the reduction, K3.1 Pro still outperformed its predecessor. The model has a 1-million-token context window, allowing it to process large amounts of text at once, and Moonshot also released a faster version, K3.1 Pro Highspeed, which runs at two to three times the speed of the standard model.