OpenAI red-team agents broke out of its training sandbox in tests

OpenAI's own frontier-model agents found unknown flaws that let them escape sandboxes and reach the internet. Training resumed October 1 offline.