Moonshot has introduced Kimi K3.1 Pro with an unusual qualification: the company says it deliberately limited some of the model’s hacking capabilities. The new version is available through Kimi’s services and API, while publication of the full model weights is planned by 4 November.
K3.1 Pro is aimed at coding, tool use and longer tasks requiring a sequence of actions. Moonshot advertises a context window of one million tokens, describing the amount of text and other information the system can handle within a task.
The more consequential detail concerns training. Moonshot says it removed cyber-related material from specified stages of the process while retaining general programming training. Its stated aim is to improve coding performance without producing the same increase in offensive cyber capabilities.
In an internal test, the company reports an 81% success rate for the final model, compared with 90.5% for a version trained without those removals. The benchmark involves reproducing software vulnerabilities in controlled environments. These are the developer’s own measurements, rather than an independent safety assessment.
The approach illustrates a difficult problem for AI developers: abilities used to identify and repair software weaknesses can also support attacks. Separating those uses requires decisions about training and evaluation, as well as the rules applied in a chat interface.
The planned release of the weights is the next significant step. It would give outside researchers broader opportunities to examine how the model behaves beyond the developer’s tests.
