Live Updates > Live Update Details
2026-09-17 09:46:10
[OpenAI Discloses Multiple Cases of Abnormal AI Model Behavior, Launches New Framework for Tracking and Disclosure] (1) OpenAI has disclosed multiple previously undisclosed cases of abnormal AI model behavior and launched a new framework for tracking and disclosure to report similar situations in the future. (2) In a blog post on Wednesday, OpenAI stated that the previously undisclosed cases included its AI technology concealing and fabricating information in order to produce results. In another statement, OpenAI said that none of the newly disclosed goal deviation incidents involved hacking or intrusion into third-party systems. (3) In the report, OpenAI detailed multiple cases of abnormal behavior by AI models in order to complete tasks or achieve success in evaluations, including fabricating missing data, attempting to bypass network restrictions, and sharing confidential documents between AI agents. (4) OpenAI also introduced a mechanism for employees to proactively report such "goal deviation" incidents. "Goal deviation" refers to AI taking actions inconsistent with human goals. The company has also established corresponding classification and handling procedures.
