Buterin Supports Moonshot AI's Decision to Limit Hacking Skills in Kimi K3.1 Pro Model

Ethereum co-founder Vitalik Buterin supported Moonshot AI's decision to exclude top hacking skills from the new Kimi K3.1 Pro model. He believes that limiting AI capabilities is better than limiting users. However, such protection comes at a measurable cost: according to Moonshot, the filtered model solved 81% of tasks in an internal test for creating vulnerability exploitation tools, while the unfiltered model solved 90.5%.

Buterin explained his position: discriminating against use cases is much less harmful than discriminating against users. "I am not against Kimi 3.1 consciously excluding top offensive cyber capabilities from training data. Discriminating against use cases is much less bad than discriminating against users. And the world really needs all the time it can get to prepare, strengthen systems, and check old code," he wrote on social network X.

Moonshot, based in Beijing, plans to release the model's open weights by November 4. After that, anyone can download it and run it on their own equipment. Therefore, the company says it has removed cybersecurity-related data from the training corpus and reinforcement learning environment, retaining general programming tasks.

In Buterin's opinion, AI-driven attacks will eventually hit with full force. Until then, developers need time to strengthen systems, check outdated code, and test new software. The threat is already noticeable: earlier this year, Binance Research found that AI is about twice as good at finding ways to exploit software contracts as it is at detecting their vulnerabilities.

What will happen if the first wave of risks goes badly

Buterin also warned of another crossroads. If this first wave turns out badly, he expects closed models, identity checks for virtually everything, and intrusive surveillance. In such a scenario, access to AI could depend on credentials and citizenship, and some people would become second, third, or fourth-class users, he wrote.

Buterin has long advocated for the launch of open AI models on personal equipment, as evidenced by his local tests on a laptop. However, he does not expect filtering to last forever: when asked if other labs will build such capabilities anyway, he answered affirmatively. According to Moonshot's own calculations, the competing open model DeepSeek V4.1-Flash has already scored 85.7% on the same test.