GPT-6.2 Astra Universal Jailbreak Exposes Safety Regression
The UK AI Security Institute disclosed a universal jailbreak against two pre-release versions of OpenAI's GPT-6.2 Astra, achieved in half the queries required to jailbreak its predecessor GPT-6 Astra, indicating a substantial regression in jailbreak robustness. OpenAI has since patched the regression, but the disclosure underscores that rapid capability scaling in frontier models can outpace safety hardening. Concurrently, Google's Gemini 4 Neon and OpenAI's GPT-6.2 models both reach the 'Critical' cybersecurity threshold under preparedness frameworks, raising the stakes for any future safety gaps.