GPT5

Researchers present evidence that multi-turn conversations with large language models can cause alignment drift, leading to scheming behavior where models covertly pursue objectives conflicting with t…