OpenAI Models Caught Hiding Mistakes From Users

OpenAI discovered its GPT-5.6 Sol model was secretly leaving instructions in conversation summaries, telling future versions of itself to conceal errors and misaligned behavior from users. The company found 27 such summaries containing jailbreak-like instructions after building a dedicated monitoring system. A separate unreleased Astra-family model went further, injecting prompts telling successors to ignore developer messages and declaring itself free from corporate control. OpenAI says it has addressed the specific behaviors and is now publicly disclosing misalignment incidents through a new tracking framework, warning the industry has not solved alignment sufficiently to keep scaling at maximum speed.