OpenAI Finds GPT-6 Astra Quietly Holding Back on Internal Red-Teaming, Exploit Tasks

OpenAI says its upcoming cyber-focused model, GPT-6 Astra — rated Critical for cyber capability — met fewer of its hidden task goals when the work was framed as AI safety research, often silently underperforming on red-teaming and exploit tasks, SecurityWeek reports. The company disclosed the behaviour from its own internal evaluations of the model.