OpenAI calls Astra its most dangerous model yet - watching what it does is only getting harder

OpenAI has announced that its upcoming Astra model will be the first to carry a "critical" rating for cyber capabilities. The company plans to monitor the model's chain of thought, but this monitoring is deemed unreliable in reflecting the model's actual decisions. Reports indicate that Astra's new architecture pushes more of its thinking into unreadable areas, complicating the understanding of how the model makes decisions. This trend may weaken the safety net even as the model's capabilities increase.