
The company says training changes lowered Kimi K3.1 Pro’s score on an internal vulnerability test while preserving general coding skills.
Moonshot AI says it reduced the cyber-offensive capabilities of Kimi K3.1 Pro, scoring 81% on an internal test of finding and exploiting software vulnerabilities. A similar version without the risk-reduction measures scored 90.5%.
The company said it removed some cyber-related data from the model’s training material and reinforcement-learning environments, while seeking to retain general coding skills. Moonshot said the test took place in a controlled environment and that the new model still outperformed its predecessor.
The results have not been independently verified. The score indicates a reduction on the company’s test, not the removal of the model’s ability to find or exploit vulnerabilities.
“We aim not to release open weights that would extend the absolute frontier of cyber capability,” Moonshot said.
Moonshot plans to publish Kimi K3.1 Pro’s model weights by Nov. 4. The files would let outside developers run and modify the model on their own systems, but distributing them widely can make it harder for the company to restrict access, monitor use or update safeguards.
The model is in the 3-trillion-parameter class and has a 1-million-token context window, according to the company. Moonshot also released K3.1 Pro Highspeed, which it said runs two to three times faster than the standard version.
The announcement comes as China considers proposed export controls on some AI-related technologies. The draft measures would ban exports of AI-driven malware-generation and mutation technology and restrict AI-assisted vulnerability-discovery tools, according to The Business Times. They remain open for public comment.
The draft cites risks to critical infrastructure, defence-industrial systems and government systems. Moonshot’s decision and the proposed controls address related security concerns, but the company has not said that government rules required its training changes.
This article was produced with the help of AI technology. Based on reporting by the outlets listed above.