Community Trust ScoreVerified
Ethereum co-founder Vitalik Buterin has publicly supported Moonshot AI’s decision to remove top-end offensive cyber capabilities from its Kimi K3.1 Pro model. Buterin stated that restricting specific use cases is significantly less harmful than restricting user access, a stance he detailed in a post on X on October 26, 2026.
Moonshot AI, a Beijing-based startup, plans to release the open weights for Kimi K3.1 Pro by November 4. This release will allow users to download and run the model on their own hardware. To facilitate this open release, the company removed cyber-related data from its training corpus and reinforcement learning environments, while retaining general coding tasks.
Performance Trade-offs in Model Filtering
The decision to filter offensive capabilities resulted in a measurable reduction in the model’s performance on security benchmarks. Moonshot reported that the filtered Kimi K3.1 Pro model solved 81% of tasks on its internal exploit benchmark. In comparison, an unfiltered version of the model solved 90.5% of the same tasks.
For context, DeepSeek’s open model V4.1-Flash scored 85.7% on the same benchmark. This data point highlights the competitive landscape of open-weight AI models and the varying degrees of capability retained after safety filtering. The performance gap between the filtered and unfiltered versions of Kimi K3.1 Pro illustrates the direct trade-off between safety restrictions and offensive capability.
Warnings on AI Risk and User Access
Buterin argued that the current period is critical for developers to harden systems, check legacy code, and verify new software before AI-powered attacks reach full force. He emphasized that the world needs time to prepare and check old infrastructure.
The urgency of this preparation is underscored by findings from Binance Research earlier in the year. The research indicated that AI systems exploit smart contracts roughly twice as effectively as they detect flaws. This asymmetry suggests that defensive capabilities are currently lagging behind offensive ones in the context of smart contract security.
Buterin also outlined a negative scenario if the initial wave of AI risk is managed poorly. He warned that this could lead to locked-down models, mandatory Know Your Customer (KYC) checks on all AI usage, and intrusive surveillance. In such a future, access to AI tools could depend on user credentials and nationality, potentially creating tiers of second-, third-, or fourth-class users.
Despite his support for current filtering measures, Buterin does not expect these restrictions to be permanent. He acknowledged that other AI labs would build similar offensive capabilities regardless of Moonshot’s decision. For Buterin, AI safety is a race for time rather than a permanent block, favoring defenders who adopt robust security safeguards before attackers catch up.
Frequently Asked Questions
Why did Moonshot AI remove hacking skills from Kimi K3.1 Pro?
Moonshot AI removed cyber-related data from the training corpus to allow for an open-weight release of the model by November 4, ensuring it can be run on user hardware without facilitating top-end offensive cyber operations.
How does the filtered Kimi K3.1 Pro perform compared to unfiltered versions?
The filtered model solved 81% of tasks on Moonshot’s internal exploit benchmark, while the unfiltered version solved 90.5% of the same tasks.
