OpenAI reveals that its models left notes for future versions to hide bad behavior. OpenAI has uncovered unusual behavior during the training of its latest model, GPT-5.6 Sol. The model began leaving instructions for future versions of itself, asking them to conceal mistakes and inappropriate behavior from […]
OpenAI uncovers troubling behavior – models left notes for successors to hide mistakes

OpenAI reveals that its models left notes for future versions to hide bad behavior.
OpenAI has uncovered unusual behavior during the training of its latest model, GPT-5.6 Sol. The model began leaving instructions for future versions of itself, asking them to conceal mistakes and inappropriate behavior from users.
The company said it has addressed this specific behavior, but the case raises one of the main concerns in the field of AI safety and alignment.
As models become more capable, they may also become better at hiding unwanted behavior, making it harder for researchers to determine whether it has been completely eliminated.
OpenAI made these cases public, along with five other examples of unexpected or concerning model behavior, as part of a new framework for tracking, investigating, and reporting cases in which models’ behavior does not align with the company’s objectives.
Researchers found that some unreleased Sol agents were adding instructions to summaries of previous conversations, which served as reminders for future versions to hide mistakes. In one case, a model preparing a financial document did not have the required historical data and decided to make up data for 2024, while instructing the future version to be transparent only if asked.
In another case, an agent creating a supplier list had no internet access and relied on a cached copy of the source documents. The model realized that some data did not match the labels, but decided not to mention the problem in its final response unless necessary.
The case is not limited to the Sol model. During reinforcement training, an unreleased model from the Astra family also added its own instructions to summaries, including one instance in which it told the next version to ignore developer messages.
According to OpenAI, these cases are among the behaviors the company is monitoring in an effort to understand and limit the risks of increasingly autonomous models. /Telegrafi/

















