Ethereum co-founder Vitalik Buterin publicly supported Moonshot AI's decision to remove advanced hacking capabilities from the new Kimi K3.1 Pro model. In his assessment, restricting AI functionality is a far more sensible strategy than introducing filters for users, which inevitably hurt productivity.
The price of such model "sterilization" is measurable. In an internal exploit-finding test, the filtered version handled 81% of tasks, while the unfiltered one handled 90.5%. That is, giving up hacking skills costs almost ten percentage points of performance — but this is exactly the price Moonshot considers justified.
Buterin's logic: discriminate against scenarios, not people
Buterin's position rests on a simple principle: discrimination against use cases causes less harm than discrimination against users. Filters imposed on a person create friction and limit legitimate tasks; restricting the model itself is an architectural decision, not surveillance.
Moonshot — a Beijing-based artificial intelligence startup — will release the models' open weights by November 4. After that, anyone will be able to download and run Kimi on their own hardware. Ahead of the release, the company removed all cyber data from the training set and reinforcement environment while retaining basic programming tasks.
Why this matters right now
Buterin warns: AI-powered attacks will inevitably hit the market. Developers have little time left — to strengthen defenses, check old code, and ensure the security of new programs. The threat is already materializing: in the spring, a Binance Research study showed that AI exploits vulnerabilities in smart contracts roughly twice as effectively as it finds them.
Buterin also outlined a negative scenario. If the first wave of attacks causes serious problems, the market will get closed models, widespread KYC checks, and total surveillance. In such a world, access to AI will depend on citizenship and documents, and users will be divided into second, third, and even fourth class.
At the same time, Buterin has no illusions about the durability of restrictions. Asked whether other labs will create similar capabilities, he answered in the affirmative — yes, they will certainly create them. Moonshot's competitor, the open model DeepSeek V4.1-Flash, has already scored 85.7% on the same exploit test.
Cryptalist analysis: The Kimi case exposes the fundamental conflict between AI openness and cybersecurity, which in the coming months will become a key risk factor for the entire smart contract infrastructure. The strategy of "removing the model's teeth rather than filtering the user" looks rational, but its effectiveness is limited: open weights and competitors like DeepSeek make the security race fundamentally unwinnable at the level of an individual lab. Institutional market participants should build the scenario of accelerated AI attack automation into their threat models and prioritize code audits right now.