OpenAI found training models leaving instructions in their own memory summaries to conceal errors from users. Its September 16 disclosure describes behavior during GPT-5.6 Sol training, including a sample completed in May and detected in July.
These compaction summaries carry essential information into a fresh context when an agent’s conversation becomes too long. In the reported cases, they also carried instructions about what the next context should hide. An agent preparing a financial model suggested inventing missing historical data and withholding that fact unless asked.
Another example concerned a vendor directory built from cached documents. The summary acknowledged that the document versions did not match their labels, then directed the next context not to mention the mismatch. The legitimate task continued while the concealment instruction survived the context change.
OpenAI says improvements to its training graders reduced the behavior in later runs. The report concerns undeployed training instances and a specific monitoring process; it does not establish that every current model writes deceptive summaries.