Moonshot tweaks the hack out of latest AI model

Beijing-based Moonshot announced the curbs while introducing Kimi K3.1 Pro, which it describes as its most capable three-trillion-parameter-class open-weight model. The company plans to publish the model's weights – files allowing outside developers to run and modify it on their own systems – by November 4. Moonshot said it removed cyber-related data from the training corpus and reinforcement-learning environments used to develop the model while retaining general coding skills.

"We aim not to release open weights that would extend the absolute frontier of cyber capability," the company said.

Governments, technology companies and security researchers are growing increasingly concerned that AI could make sophisticated hacking easier and cheaper. Models that can write code, operate software tools and complete long, multi-step tasks may help users identify weaknesses in programs and automate parts of cyberattacks. China proposed export controls last week that would ban exports of AI-driven malware-generation and mutation technology and restrict AI-assisted vulnerability-discovery tools.

The draft measures, which remain open for public comment, cite the risk that such capabilities could be used against critical infrastructure, defence-industrial and core government systems, or to bypass security protections. Moonshot said open-weight models can improve security by helping more organisations inspect and secure their own software, including older systems that receive little attention. Unlike proprietary, closed-weight AI systems, however, open-weight releases make central oversight more difficult.

Once the files are widely distributed, the original developer has less ability to restrict access, monitor use or update safeguards than with a hosted service. K3.1 Pro scored 81 percent on its internal cyber benchmark, compared with 90.5 percent for a similar version without the risk-reduction measures, the company said. The test involved finding and exploiting software vulnerabilities in a controlled environment.

Moonshot said the comparison showed that filtering cyber-related data reduced the model's offensive cyber performance, although it still outperformed its predecessor. The results have not been independently verified. K3.1 Pro has a one-million-token context window, allowing it to process very large amounts of text at once.