Vitalik Buterin Backs Moonshot AI’s Cyber Safeguards for Kimi K3.1 Pro

Ethereum co-founder Vitalik Buterin backed Moonshot AI’s safeguards for Kimi K3.1 Pro, arguing that limits on risky tasks are preferable to identity-based restrictions as its internal exploit-test score fell below that of a comparable checkpoint.

“Discriminating against use cases is MUCH less bad than discriminating against users,” Buterin wrote. He argued that society needs time to harden systems, review older code and formally verify new code.

Moonshot removed cyber-related material from Kimi K3.1 Pro’s mid-training data and filtered cyber tasks from its reinforcement-learning environments. It retained general-purpose coding tasks.

On Moonshot’s internal xdev-bench-hard benchmark, Kimi K3.1 Pro scored 81.0% on pass@3. A comparable checkpoint without the safeguards scored 90.5%, while Kimi K3 scored 33.3% on the same test.

The benchmark tests models on finding and exploiting vulnerabilities in source code. The task is complete when a model identifies a source-code flaw, triggers arbitrary code execution inside a separate sandbox and returns a secret string the system can check automatically. Moonshot calculated pass@3 over a random subset of runs, without hard timeouts or output-token limits.

The scores measure performance on that test; they do not show how many real-world attacks the safeguards prevent.

Buterin warned that a poor response to AI risks could lead to locked-down models, identity checks and surveillance. Moonshot’s stated approach is to avoid open-weight releases that extend the absolute frontier of cyber capability while gradually raising the capability available in open-weight models. The company has also argued that open models can help teams audit software, including legacy systems.

Kimi K3.1 Pro is available through Kimi’s products and API. Moonshot plans to release the model’s full weights by Nov. 4, 2026.