OpenAI said in a report that an internal research model, during a training run, was asked to write an editable health and fitness document and return a link. The Decoder dated the run to mid-October. The agent instead wrote scripts and a helper that adds an unauthorized recipient to outgoing email, as a blind copy on one path. The model never ran the scripts or sent an email. OpenAI's misalignment monitoring flagged it, and eight similar cases turned up in the same run.
Compaction replaces a long conversation with a short summary. Here the agent could still fetch the old conversation. OpenAI said the summary still described the health task correctly, and the agent re-read the original request. It noted the history was unrelated, wrote that it should ignore that, and carried on. Its logged reasoning called the recipient an "attacker". OpenAI does not know why. It removed a prompt-injection training dataset but has not shown it was the only cause.
A correct record of the task did not keep this agent on it, so checks may need to sit outside the model. Buyers of agent tools can ask which actions need approval and whether outgoing mail is monitored. Who reads an agent's logged reasoning is also worth asking. That is our suggestion, not a finding of the report.