Vitalik Buterin backed Moonshot AI’s decision to limit Kimi K3.1 Pro’s offensive cyber capabilities. Kimi’s cybersecurity benchmark score fell from 90.5% to 81% after restrictions. Ethereum c
Ethereum co-founder Vitalik Buterin has backed Moonshot AI’s decision to limit the offensive cybersecurity capabilities of its latest Kimi K3.1 Pro model, arguing that reducing dangerous AI functions is preferable to restricting access based on users’ identities.
Buterin shared his position on October 26, following Moonshot’s announcement that it had removed cyber-related material from the model’s training data and reinforcement-learning environments. The Chinese AI startup plans to release the model’s weights on November 4, allowing developers to download, run and modify it on their own systems.
I am okay with Kimi 3.1 deliberately leaving out top-end offensive cyber capabilities from their training data. Discriminating against use cases is MUCH less bad than discriminating against users. And the world does need all the time it can get to prepare, harden, check old code, (properly) formally verify new code, etc before the full weight of very powerful cyber attacks is upon us. And if this first big wave of AI risk is not handled well, then we may well get a dystopia of highly locked down models, KYC gating on everything, highly intrusive surveillance on everything, a society where you may be a second or third or fourth-class citizen in terms of model access based on your credentials and nationality.
— vitalik.eth (@VitalikButerin) October 26, 2026
In his post on X, Buterin said he was comfortable with Kimi deliberately leaving out advanced offensive cyber capabilities. He argued that restricting specific uses of AI is less harmful than limiting who can access the technology.
Buterin also warned that society needs time to strengthen existing software, examine older systems and use formal verification to improve the security of new code before more powerful AI-driven cyberattacks become widespread.
Moonshot said the changes reduced the model’s performance on its internal cybersecurity benchmark while preserving its general coding abilities.
The restricted version scored 81% on the test, compared with 90.5% for a similar version without the safeguards. The difference is 9.5 percentage points. The benchmark involved finding and exploiting software vulnerabilities in a controlled environment, and the results have not been independently verified.
Despite the reduction, Kimi K3.1 Pro still scored considerably higher than its predecessor, Kimi K3, which achieved 33.3% on the same assessment, according to Moonshot.
The decision comes as governments and cybersecurity researchers warn that increasingly capable AI tools could make identifying software weaknesses and developing exploits faster and cheaper. China has also proposed export controls targeting AI-driven malware generation and certain AI-assisted vulnerability-discovery technologies.
Open-weight models present an additional challenge because developers can distribute and modify them outside the original provider’s control. Moonshot argues that such models can also help security teams examine vulnerable software, including older systems that receive limited maintenance.
Buterin’s comments highlight the challenge facing AI developers: reducing the potential for harmful cyber activity without imposing broad restrictions on legitimate users. Moonshot’s planned November 4 release will put its approach under closer scrutiny as developers and researchers gain wider access to the model.
