Kimi K3.1 Pro Cyber Curbs: 81.0% vs 90.5% [2026]

Moonshot AI says it dialed back cyber-related learning in Kimi K3.1 Pro to reduce misuse before release. Citing an internal test, Reuters and The Information report the model scored 81.0% at finding and exploiting software vulnerabilities, while a similar version built without the risk-reduction measures scored 90.5%. The Beijing startup positions this as misuse reduction by removing cyber material both from the training corpus and from reinforcement-learning environments, rather than adding a refusal filter to a finished model. The company also argues that it aims not to release open weights that would extend the “absolute frontier” of cyber capability. The results are not independently verified, and the article says the difference matters as security teams wait for outside testing.