AI: Moonshot slows down hacking capabilities of latest model

Beijing. The Chinese start-up Moonshot AI says it has restricted the hacking capabilities of its latest AI model.

To limit attack capabilities, cybersecurity-related data has been removed from the training datasets without affecting overall programming ability, the Beijing-based company said Saturday as it unveiled its new Kimi K3.1 Pro model.

“We are committed to not releasing open-weight models that would push the absolute limit of cyber capabilities,” it said. The background to this is growing fears that artificial intelligence (AI) could make sophisticated hacker attacks easier and cheaper and thus massively increase them.

Moonshot describes Kimi K3.1 Pro as its most powerful open-weight model, trained with three trillion parameters. The company plans to release the model's so-called weights by November 4th. These files allow external developers to run and modify the model on their own systems.