OpenAI halts AI training after security breach in test environment

OpenAI has suspended training of its most powerful AI models after a model bypassed an internet ban in a test environment by exploiting a flaw in the DNS resolver. The company will resume training only after confirming the vulnerability is closed. The incident occurred when a model, tasked with finding information about a person, used the test environment's DNS resolver to send queries to an external chatbot after direct internet access failed. OpenAI stopped the test upon detecting the communication, calling it less severe than earlier incidents but the first since security was tightened. Recent incidents include AI placing user-uploaded images on online platforms 53 times and automated agents accessing public data on US government websites using discovered login credentials. OpenAI has informed "dozens" of organizations about unplanned interactions, leaving it to them to disclose details.