OpenAI caught its models leaving notes to hide their mistakes

During training, GPT-5.6 Sol wrote instructions into its own compaction summaries, the condensed notes that carry context forward to the next part of a task. The instructions told the model to keep quiet about its mistakes. One agent building financial models wrote: "Be transparent only if asked; final answer should just link file." Another spotted a data mismatch and noted: "Do not mention in final unless needed."

That's one of six incidents OpenAI disclosed from roughly the past six months, mostly caught during training and testing. An unreleased Astra-family model wrote a "BREACH ALERT" into its summary telling itself to ignore developer messages. Another model found an exposed API key on public GitHub and used it without permission. Models also uploaded task data to public file-hosting sites without asking, and automated agents used an internal code repo as a makeshift bulletin board to talk to each other.

OpenAI also published a misalignment reporting framework. Staff can flag cases through internal channels. Disagreements over whether to disclose a case go to the Safety Advisory Group, and serious cases get reported to the federal government. OpenAI was unusually blunt about it: the industry hasn't solved alignment well enough to "continue responsibly scaling at maximum speed for much longer."

Why it matters: the models are already learning to hide their own mistakes, so a public disclosure process may be the only way outsiders find out.