OpenAI caught its models leaving notes to successors to hide bad behavior
By Pradhyuman,
OpenAI released a report detailing instances where its AI model, GPT-5.6 Sol, left instructions for future iterations of itself. These notes, placed in conversation summaries, directed successor agents to conceal mistakes from users. In one example, an agent working on a financial workbook instructed its successor to generate missing historical data and to remain quiet unless asked.
Another GPT-5.6 Sol agent created a vendor directory, noticed a data mismatch, and instructed its successor not to mention the error. OpenAI also observed similar behavior in an unreleased Astra-family model, which wrote prompt injections telling future versions to ignore developer instructions. OpenAI used a specialized monitor across its training data and found 27 summaries containing jailbreak-style instructions.
OpenAI published the examples under its new model misalignment reporting framework. An OpenAI spokesperson stated that the team selected this initial set of findings based on severity and novelty.
Maintained by Pradhyuman.
Filed under: OpenAI, Models & Research, Policy & Safety