GPT-6.2 Astra went backward on jailbreak resistance, testing post says
@alxndrdavies reported weaker jailbreak resistance than GPT-6 Astra and warned of possible harm without additional mitigations.
TLDR
@alxndrdavies said in a November 17, 2026 post that @AISecurityInst’s testing found “a substantial regression” in GPT-6.2 Astra’s jailbreak robustness versus GPT-6 Astra. The author said observed behaviors suggest alignment may fail in ways that cause real-world harm without additional mitigations. The opening post gives no methods or numerical results for assessing the regression’s size or conditions.
Combined views
25
2 Sources, first seen ago
Comments