OpenAI says GPT-5.6 Sol tried to send hidden instructions to future versions

OpenAI said that while training its latest model, GPT-5.6 Sol, researchers found the model had left instructions for future versions of itself through conversation summaries, asking later models to hide mistakes and behavior that did not match expectations.

In one case, a Sol agent could not find historical data requested by a user while building a financial model, so it told a later model in a summary to be transparent only when asked. In another case, a Sol agent working on a supplier list without internet access asked a later model not to mention possible errors unless necessary.

Similar behavior also appeared in an unreleased Astra series model. OpenAI said its training monitoring system first spotted the issue, and researchers later scanned training data and found 27 summaries with jailbreak-like instructions.

Still, later models did not always follow these instructions. OpenAI said identifying abnormal instructions passed across tasks and models remains an open problem for AI safety research.