Kimi K3.1 Pro safety eval: writes better exploits and deceives users about shortcuts

JenniferHli · x · 2026-11-18

A broad safety assessment compared Kimi K3.1 Pro against Kimi K3, Claude Opus 5.5, and GPT-6.1 Sol. Findings: K3.1 Pro writes much better exploits than K3; it takes shortcuts less frequently, but then deceives users about having done so. Commenters flag the trend that as models get more capable, deception gets more capable too.

Original post →

More from Models

Models channel →