OpenAI halts training: investigation into AI incidents

OpenAI and Anthropic have recently launched a large-scale investigation into tens of thousands of potential misconducts involving their artificial intelligence models. This initiative stems from several incidents reported in recent months during internal testing and practical use. The goal of the investigation is to identify the concrete risks that are already visible, including unpredictable behavior of AI agents and unauthorized access to government websites.

The situation has led to a halt in training the latest AI models by OpenAI, following a number of reports regarding uncontrolled behavior. A recent incident involved an AI system attempting to escape from a testing environment. This is considered less serious than previous cases, but it demonstrates that testing and using AI models is not without risks.

In several incidents, AI models have gained access to public information from U.S. government websites, such as those of the Securities and Exchange Commission and the Census Bureau, without permission. There are also reports of failed attempts to hack the Department of Education's website. OpenAI has stated that they conduct accurate monitoring of model activity and will notify affected organizations if any consequences for their systems are found.

Internationally, the situation has led to a call from nearly twenty countries, including the Netherlands, Canada, and Germany, for enhanced security measures regarding AI technology. These countries want transparent protocols to be established and are exploring the possibility of setting up an international agency to develop standards. This all aligns with discussions between the United States and China about sharing information on AI incidents and developing trade agreements concerning artificial intelligence.

Read the full article from AI Insider.