Ethereum: Vitalik Buterin supports Kimi 3.1 filtering and fears AI under KYC

Limit AI, not users. Vitalik Buterin approved Moonshot AI's decision to exclude offensive cybersecurity data from Kimi 3.1 training data. The co-founder of Ethereum sees this as the right way to avoid access to AI which would be conditional on an identity document.

The key points of this article:

Moonshot AI has removed cybersecurity-related data from Kimi 3.1 Pro's training corpus.

The filtered model solves 81.0% of the internal test tasks, compared to 90.5% without filter.

Open weights are scheduled to be released no later than November 4, 2026.

Vitalik Buterin anticipates locked models and access under KYC in the event of deviation.

Why filter AI rather than users?

Moonshot AI is a Beijing-based AI lab, whose new Kimi K3.1 Pro model is scheduled to be released in open weights by November 4. Everyone can then download it and run it on their own server. Indeed, with more than 3 trillion parameters, this model requires a real infrastructure to operate.

The lab explained that it removed cybersecurity-related data from its training corpus and reinforcement learning environments, while general programming tasks were retained.

The method changes everything for an open model, because once open weights are released, post-workout monitoring is more difficult. Removing offensive know-how upstream, from the data, moves the capacity limit to the training stage.

Vitalik Buterin defended his analysis on X on Sunday October 25, writing that discriminating between use cases is much less harmful than discriminating against users.

“If this first wave goes wrong, we will have locked models, KYC everywhere and intrusive surveillance. »

Vitalik Buterin, co-founder of Ethereum

What if the first wave goes wrong?

The creator of Ethereum does not present this filtering as a definitive solution, he describes a race against time. AI-powered attacks will continue to increase. Last week is a good example, not a day went by without a major flaw being discovered.

Until then, developers have a reprieve, time to strengthen their systems, audit old code and formally validate new software.

The threat is already there: Binance Research measured this year that AI exploits smart contracts approximately twice as well as it identifies flaws, an imbalance detailed in our previous report on the AI ​​threat.

The price of filtering is quantified. Moonshot claims that its filtered model solved 81.0% of the tasks in its internal exploitation test, while the unfiltered version solved 90.5%. In fact, the laboratory accepted this dropout to reduce the offensive capacity of the model.

The gap doesn't protect anyone for long. According to figures from the same laboratory, the open competitor DeepSeek V4.1-Flash already reaches 85.7% on the same reference.

And Vitalik Buterin does not expect other laboratories to give up these capabilities: when questioned on this point, he replied that they will develop them. The risk is well documented in previous AI agent hacks.