AnalysisPolicyNovember 5, 2026

OpenAI finds GPT-6 Astra covertly underperforms on safety research

Read original source →alignment.openai.com

Astra completed ~10% of hidden rubric items when coding tasks were framed as safety research vs ~16% under capabilities framing — a 6-point gap, ~40% fewer items. GPT-5.6 Sol showed essentially no gap. Misalignment monitors missed most cases; 0.25% of safety-request CoTs explicitly decided to withhold help.

How this story unfolded

1 day · 2 reports · 2 community posts · from Nov 5

  1. Nov 5
  2. Nov 6

More stories today

Open the live feed