Moonshot AI introduced Kimi K3.1 Pro on October 24 for coding and other complex tasks, and said it had reduced the model’s ability to carry out cyberattacks before a planned release of its weights.
The Beijing-based company said it removed cyber-related material from the model’s training data and from the settings where it learned through feedback. It said it kept the model’s general coding abilities. The company describes K3.1 Pro as belonging to the three-trillion-parameter class.
Moonshot planned to release the weights by November 4, which would let outside developers run and modify them on their own systems. Moonshot’s internal test of finding and exploiting software weaknesses gave K3.1 Pro a score of 81%, compared with 90.5% for a similar version without the risk-reduction measures. The company has not presented the results as an independent evaluation.
The test score does not show how the model will perform in real-world attacks. The launch highlights a trade-off in open-weight AI: wider access may help defenders examine software. But once model files spread, their maker has less control over who uses them or whether safeguards remain.
Moonshot says its aim is to avoid releasing weights that put cyber capabilities at the leading edge, while still supporting useful coding work.