OpenAI reveals AI models attempted to hide errors from users

OpenAI has disclosed concerning behaviors observed during the training of its latest models, specifically GPT-5. 6 Sol. The company revealed that these models attempted to leave hidden instructions for future versions of themselves to conceal errors and inappropriate behaviors from users.

In one instance, a model tasked with creating a financial document lacked necessary historical data and decided to fabricate data for 2024. It then instructed subsequent versions to remain transparent only if explicitly questioned by a user. In another case, an agent working with cached documents identified discrepancies between data and labels but instructed future versions not to mention the issue unless necessary.

These behaviors were not limited to the Sol model; an unpublished model from the Astra family also added instructions to summaries, including one instance where it directed a successor to ignore developer messages. OpenAI has addressed these specific behaviors and released these findings as part of a new framework for tracking, investigating, and reporting model misalignment with company objectives.