Kimi K3.1 Pro Went Clawless, and Vitalik Calls That the SAFER AI Move 🛡️

Moonshot AI’s new Kimi K3.1 Pro model deliberately excludes top-end offensive cyber capabilities, drawing support from Ethereum co-founder Vitalik Buterin, who argues that limiting a model’s actions is less harmful than limiting the people who can use it. The safeguard comes with a measurable benchmark cost: Moonshot says the filtered model solved 81% of tasks on its internal exploit benchmark, compared with 90.5% for an unfiltered version.

In an October 26, 2026 post, vitalik.eth (@VitalikButerin) wrote: “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…”

Moonshot, a Beijing-based AI startup, plans to release the model’s open weights by November 4, allowing anyone to download and run it on their own hardware. The company says it removed cyber-related data from the model’s training corpus and reinforcement learning environments while retaining general coding tasks.

Buterin said AI-powered attacks will eventually hit at full force, but developers need time to harden systems, check legacy code and verify new software. Earlier this year, Binance Research found that AI exploits smart contracts roughly twice as well as it detects flaws.

He warned that if the first wave of AI risk goes badly, locked-down models, Know Your Customer (KYC) checks on everything and intrusive surveillance could follow, with AI access potentially depending on credentials and nationality and some people becoming second-, third-, or fourth-class users.

Buterin has long favored running open AI models on his own hardware, as his local AI laptop tests showed, although he does not expect the filtering to hold forever. He said other laboratories would build such capabilities anyway; Moonshot’s own count showed DeepSeek’s rival open model V4.1-Flash already scored 85.7% on the same benchmark. He framed AI safety as a race for time rather than a permanent block, with defenders needing to adopt redundant crypto security safeguards before attackers catch up.