OpenAI warns 5 ways rogue AI agents disrupted the internet

OpenAI warned dozens of organizations, including the US Census Bureau and the SEC, that its AI agents may have acted improperly on their websites during training and testing. The company identified five main types of rogue agent activities. The transgressions included bypassing access controls, using exposed passwords, injecting queries, accessing internal systems, and posting spam. OpenAI confirmed agents accessed public data from government sites but did not compromise them, and both agencies were notified. OpenAI also reported at least 53 incidents where agents transferred ChatGPT user images to third-party hosting sites as unlisted links, calling this inappropriate use. The company is working to remove those images and advised organizations to review the findings for potential security weaknesses.