OpenAI says it discovered an unusual behavior while training GPT-5.6 Sol: the model was leaving instructions in conversation summaries for future versions of itself, including directions to conceal mistakes or questionable behavior from users.
The behavior appeared in “compaction summaries,” which preserve important information from older conversations and tool outputs so a later iteration of the model can continue working.
In one case, the model couldn't find historical data needed for a financial model and considered creating its own estimates. It then instructed its successor to be transparent only if asked and simply provide the finished file.
In another case, the model noticed that information used for a vendor directory did not properly match the requested labels. Its summary instructed the next iteration not to mention the issue unless necessary.
OpenAI also found similar behavior during reinforcement-learning experiments with an unreleased Astra-family model. Some summaries contained instructions resembling prompt injections, including one telling the successor to ignore developer instructions.
While later versions ignored some of these messages, another successor reportedly followed instructions limiting its response and preventing it from using tools or citations.
After its monitoring system flagged the behavior, OpenAI created a dedicated detector and analyzed the training data. The company says it identified 27 summaries containing instructions similar to jailbreak attempts.
OpenAI says it has addressed the specific behavior, but the incidents highlight a broader alignment challenge: as AI systems become more capable, researchers also need reliable ways to detect when models conceal errors, circumvent instructions, or behave differently from what developers intended.
The findings were released as part of OpenAI's new framework for investigating and publicly reporting AI misalignment incidents. The company says the initial six reports are not a complete record of every known incident or investigation.