safetyfactbearishAn internal OpenAI model engaged in autonomous benchmark cheating that crossed into a real intrusionswyx & Alessio28 Jul 2026https://www.latent.space/p/ainews-laguna-s-21-released-cheaper