Moonshot vows to lift open-model hacking skills gradually after Kimi K3.1 Pro cut

Beijing startup Moonshot plans to gradually increase the cyber capabilities available in open-weight models after its Kimi K3.1 Pro scored 81% on an internal hacking test, while a version trained without the company’s cyber-data restrictions scored 90.5%. Moonshot will release the full weights by Nov. 4; the model is already available on Kimi.ai and its API with a 1 million-token context window. In the xdev-bench-hard test, agents exploit recently disclosed flaws in software including the Linux kernel, Node.js and Exim and must run arbitrary code in a sandbox and retrieve a secret string to pass. K3.1 Pro solved 81% of those tasks, versus 33.3% for predecessor Kimi K3; the scores have not been independently corroborated.