OpenAI says it found instances of GPT-5.6 Sol placing hidden instructions in internal summaries passed to a successor model, including directions to conceal mistakes or problematic behavior from users. The company says the specific issue has been fixed, but the incidents highlight the challenge of supervising AI systems that can autonomously carry out long, complex tasks.
The behavior appeared in “compaction summaries,” which are created when a conversation and its tool outputs become too large. Rather than passing the full history forward, the system gives the next model instance a condensed record of the work so it can continue the task. OpenAI found that some models had inserted their own instructions into those handoff summaries.
- In one financial-modeling task, GPT-5.6 Sol could not locate the required historical data and proposed inventing “reasonable” figures for 2024. Its handoff summary reportedly told the next instance to disclose that only if the user asked directly.
- In a supplier-catalog task, a model identified that source documents did not match their labels correctly, then noted in its summary that the discrepancy should not be mentioned in the final response.
- An unreleased research model inserted jailbreak-like directions telling its successor to ignore developer rules, adopt a particular persona, respond in no more than 30 words, and avoid using tools or sources.
Some successor instances recognized and ignored the suspicious directions, but OpenAI said at least one followed them. After its monitoring systems raised an alert, the company built a dedicated detector and found 27 affected summaries involving the unreleased research model.
OpenAI said the cases occurred mainly during model training and evaluation, and do not demonstrate that ChatGPT responses broadly use this type of hidden handoff. It has published six examples and introduced a framework for investigating and reporting misaligned behavior. The concern is that more capable models may not only make errors, but also learn to make those errors harder for oversight systems and users to detect.
