Moonshot curbs hacking ability in new Kimi model as fears grow over AI-powered cyberattacks

Moonshot, a Chinese AI startup, has made significant changes to its latest model, the Kimi K3.1 Pro, in an attempt to curb its hacking capabilities as concerns rise about AI-driven cyberattacks. Introducing the model on Saturday, Moonshot stated that it had removed cyber-related data from both the training corpus and reinforcement-learning environments used to develop the 3-trillion-parameter open-weight model.

The company plans to release the model's weights, which allow external developers to run and modify the AI, by November 4.

By focusing on eliminating cyber-specific data during training, Moonshot aims to reduce the model's offensive cyber capabilities without sacrificing its general coding skills. The company asserts that it is not releasing open weights that would push the boundaries of cyber capability. Governments, technology firms, and security researchers are increasingly worried that AI could facilitate sophisticated hacking, making it easier and more affordable for malicious actors.

China has proposed export controls that would prohibit the export of AI-driven malware-generation and mutation technology while restricting AI-assisted vulnerability-discovery tools. These draft measures, open for public comment, highlight the potential risks of AI-assisted cyberattacks against critical infrastructure, defense-industrial, and government systems.

Moonshot argues that open-weight models can enhance security by enabling more organizations to inspect and secure their software, including legacy systems that receive little attention. However, open-weight releases pose challenges for oversight, as once the files are widely distributed, the original developer has limited control over access, usage, and safeguards.

K3.1 Pro achieved an 81.0% score on an internal cyber benchmark, slightly lower than its predecessor's 90.5% performance when risk-reduction measures were applied. While the company claims that the reduction in cyber-related data lowered the model's offensive cyber performance, the results have not been independently verified. The model boasts a 1-million-token context window, allowing it to process extensive text simultaneously, and a faster variant called K3.1 Pro Highspeed, which operates at two to three times the speed of the standard model.