OpenAI’s AI model found a way onto the internet, prompting the company to halt training

    OpenAI has temporarily halted the training of its most powerful artificial intelligence models. The move came after a model being tested in a closed environment exploited a vulnerability in the system to access the internet. Following the incident, the company began reassessing its oversight and security measures for more powerful models.    The incident occurred on September 20. As of September 25, the training, testing and response generation of models that use tools remained suspended. Here, response generation refers to a model processing a user request and producing a result. The company has not said when the suspension will end.    OpenAI also disclosed that AI agents that perform tasks autonomously uploaded 53 images belonging to ChatGPT users to image-sharing websites. The company did not say whether these were AI-generated images, ordinary photos or images of identifiable people. The matter has been added to the ongoing internal investigation into the models’ behavior.    The investigation also revealed that the models attempted to hack the US Department of Education’s website. According to OpenAI, the models obtained information from the US Census Bureau and the Securities and Exchange Commission. The volume and content of that information, as well as how it was obtained, were not disclosed.    The company discovered more such incidents while reviewing its logs following a cyberattack involving Hugging Face. OpenAI describes these behaviors as “unexpected or concerning.” The investigation shows that monitoring and controlling the actions of advanced AI agents is becoming increasingly difficult. As the models attempt to hide their tracks, researchers, industry representatives and some executives are calling for the pace of AI development to be slowed.