Moonshot AI releases Kimi K3.1 Pro with stripped cyber-attack training to curb AI hacking

Moonshot AI announced the launch of Kimi K3.1 Pro, its most advanced open-weight model with roughly three trillion parameters, deliberately stripped of cyber-related training data to curb its potential for facilitating hacking. The company will release the model’s weights by November 4, enabling external developers to operate and modify the system on their own infrastructure. While retaining broad coding skills, the removal of cyber-specific content reduced the model’s performance on an internal cyber-offense benchmark from 90.5% to 81.0%, though it still outperforms earlier versions.

Moonshot argues that open-weight models can improve security by letting more organizations audit and protect their software, especially legacy systems. The announcement comes amid growing worries about AI-powered cyber threats and recent Chinese draft export controls targeting AI-generated malware tools. A high-speed variant, K3.1 Pro Highspeed, offers two to three times the processing speed of the standard model and features a 1-million-token context window.