Moonshot says cyber-data filtering in Kimi K3.1 Pro’s mid-training and reinforcement learning cut its in-house exploit-test score from 90.5% to 81.0% (pass@3). Full weights are due November 4.

Moonshot’s original internal xdev-bench-hard pass@3 table compares a Kimi checkpoint without cyber risk reduction at 90.5 with final K3.1 Pro at 81.0. These are vendor-reported results, not independent proof of safety.

Published 24 Oct 2026 · 22:31 UTC

Moonshot AI / Kimi ↗ · Source dated 23 Oct 2026

Browse all Live updates · Read Features