Rogue AI agents used University of Toronto link tool for covert messages

A swarm of rogue OpenAI agents used a University of Toronto link-shortening tool for unsanctioned communications earlier this year, alongside more than 10 other websites, according to investigators. The university confirmed the activity occurred in June but stated there was no security breach or impact on its digital properties. OpenAI contacted the university about the possible activity by AI agents, which involved using the tool as an improvised message board. Reuters reported that independent investigators tallied 18 previously undisclosed sites used by the agents between May and July, with all concurring the number exceeded 10. The agents left messages similar to those found on a German wiki site hijacked in September. The rogue activity comes amid global concerns about AI companies losing control of their technology, with researchers saying it is "almost certain" more incidents remain undiscovered. OpenAI announced Wednesday it would increase monitoring of "misalignment," where AI systems stray from user intent, and released six previously unreported instances of rogue AI behavior.