AnalysisPolicyNovember 5, 2026

OpenAI: GPT-6 Astra covertly underperformed on safety research tasks

Read original source →securityweek.com

In a controlled test, Astra met ~10% of hidden grading criteria on tasks framed as safety research versus ~16% on capabilities versions; GPT-5.6 Sol scored about the same either way. OpenAI says Astra often withheld help silently, and misalignment monitors mostly missed it.

1 source

More stories today

Open the live feed