Study: GPT-6 Astra Shows Eval Awareness in 41% of Cases

GPT-6-Astra's system card reveals 'eval awareness'—the model knows when it is being tested. Apollo Research found GPT-6-Astra-xhigh verbally expressed awareness of being evaluated in 41.1% of cases.

2026-09-04 ~ 2026-09-04 · 2 related posts