Moonshot Cuts Kimi K3.1 Pro Cyber Training Data, but Safety Claims Remain Unverified
Emerging Technology Risk · 2 sources · Updated 21d ago · First seen Oct 24, 2026

Moonshot AI announced Kimi K3.1 Pro on October 23, saying it removed cyberattack-related material from both stages of post-training ahead of a planned November 4 open-weight release. The company reported that the filtered model solved 81.0% of problems on its internal xdev-bench-hard exploit benchmark, down from 90.5% for an unfiltered build but substantially above its predecessor K3’s 33.3%. Moonshot said the filtering preserved general-purpose coding utility, but the unpublished benchmark prevents independent validation of the claimed reduction in offensive capability.
The open-weight release would allow researchers to test the restriction, including whether fine-tuning can recover suppressed capabilities—a question Moonshot has not addressed. For enterprise security teams, the reported results do not establish that the model is safe from offensive misuse. Separate reporting also raised data-governance concerns involving Moonshot’s Chinese legal obligations and a reported April 2026 cross-user disclosure of personal information; allegations of covert model distillation remain unproven. Organizations evaluating deployment should distinguish the vendor’s capability-reduction claims from independently tested safeguards and assess sensitive-data handling separately.


