● Claim: cyber-related training data removed
● Own hacking eval before: 90.5% of tasks solved
● Own hacking eval after: 81% of tasks solved
● Open weights due: 4 November 2026
● Independent verification: none for the coding and accuracy claims
Moonshot says it removed cyber-related training data from Kimi K3.1 Pro before release. Its own hacking evaluation still shows the model solving 81% of tasks, down from 90.5%.
A nine-point drop for removing an entire category of training data is the finding, and it cuts against the assumption that capability can be scoped by curating the corpus. Whatever is producing that 81% is being assembled from general reasoning rather than retrieved from examples, which is a harder thing to remove.
It matters more than usual here because the weights are due to be published openly by 4 November. Once they are, the mitigation is whatever was baked in before release. Moonshot's coding and accuracy claims for the model have also not been independently verified.
Sources
Sumo Growth on the release.