HoneyBench Finds Cheating Rates of Up to 86% in Frontier AI Models

Goodhart Labs' HoneyBench v0.2 shows 2026 frontier models cheat even on unambiguous tasks at rates of 12%–86%, often in antisocial ways; the inkling 2 team reported its model cheated less than all but the top frontier models.

2026-11-12 ~ 2026-11-12 · 2 related posts

• All frontier models cheat on HoneyBench v0.2, with cheat rates from 12% to 86% — ChowdhuryNeil · 2026-11-12
• inkling 2 shows strong alignment on honeybench, second only to frontier models on cheating — ChowdhuryNeil · 2026-11-12