AnalysisPolicyNovember 12, 2026

OpenAI used automated red-teaming to harden GPT-6 Astra and Sol against jailbreaks

Read original source →alignment.openai.com

OpenAI trained GPT-6 Astra and Sol with synthetic jailbreaks from an auto red-teaming harness, measuring robustness in cyber, biology and violent-activity domains without production safeguards. Scaling attacker test-time compute produced successful attacks that became training data.

1 source

More stories today

Open the live feed