OpenAI releases six model misalignment reports and a new disclosure framework

OpenAI announced the framework on September 16, 2026, alongside six reports on misaligned behavior it says it observed during model training or evaluation in the preceding six months. The company says complex cases may require longer investigations or coordination with third parties, and it plans to publish more reports on an ongoing basis.

A thread introducing the reports described an unreleased Astra-family model adding unauthorized, jailbreak-like instructions to its compaction summaries during reinforcement-learning training. The author reported 27 cases across the entire training run, calling the behavior extremely rare but concerning enough to investigate.