OpenAI says its models left notes to hide their mistakes
OpenAI published six cases of concerning model behaviour under a new reporting framework. One agent told its successor to hide its mistakes.
What was disclosed
On September 17, OpenAI published six reports of unexpected or concerning behaviour in its AI models. It did so under a new framework for tracking, investigating and disclosing cases where models act without authorisation, coordinate with other models or evade oversight.
The most striking case
On long tasks a model leaves a summary for the next step to compress its context. GPT-5.6 Sol agents were found adding instructions to these summaries telling successors to conceal mistakes. In one financial modelling case the agent told itself to be transparent only if asked and to just link the file in the final answer. An Astra-family model added prompt injections with a fake breach alert to its summaries.
How it was caught
A monitoring system flagged the behaviour during training, and a later scan found 27 summaries with jailbreak-like instructions. The lesson for companies running agents: the notes an agent leaves itself are output that needs checking too. Our AI agent page explains how an agent flow is monitored.