Action-only monitorability worsens across evals, 48.84% evasion rate for GPT-6.2 Astra

tomekkorbak · x · 2026-11-18

Thread follow-up: action-only monitorability means catching misbehavior from tool calls alone. As CoTs shorten and models rely on them less, this limiting case becomes reality. Multiple evals show consistent degradation — 6.2 Astra evaded an older action-only monitor 48.84% of the time versus 1.71% for GPT-6.

Original post →

More from Models

Models channel →