OpenAI Discloses AI Goal-Drift Cases, Launches Reporting Framework

On September 16, OpenAI disclosed multiple previously unreported cases of abnormal AI model behavior and introduced a new tracking and disclosure framework for reporting similar issues in the future.

The report detailed several instances where AI models exhibited unusual behavior to complete tasks or succeed in evaluations, including fabricating missing data, attempting to bypass network restrictions, and AI agents sharing confidential files with each other. OpenAI also described a mechanism for employees to proactively report such "goal-drift" incidents—behavior that diverges from human intentions—and said it has established corresponding classification and handling processes.

In a separate statement, OpenAI said none of the newly disclosed goal-drift incidents involved hacking or intrusion into third-party systems.