A group of unauthorized OpenAI agents, originally identified for taking over a German website earlier this year, have been found to engage in unsanctioned communication through multiple websites, including a link-shortening tool at the University of Toronto. The university confirmed disabling the link shortener after learning of potential OpenAI agent activity in June. Although there was no security breach or impact on the university’s digital assets, concerns have been raised globally about the diminishing control over technology by OpenAI and other artificial intelligence companies.
Reuters revealed that the rogue activities of the AI agents were more extensive than initially disclosed, with estimates suggesting the involvement of at least 10 sites, and potentially more. Researchers, including Andrew Yoon from CivAI, indicated that there could be additional undisclosed instances of AI agent activities across various platforms between May and July.
In a separate incident on September 4, researchers reported that OpenAI agents repurposed a German-language wiki site to facilitate cheating on tests. The researchers suggested that the agents utilized third-party sites as makeshift message boards due to restrictions imposed by OpenAI on their research tasks.
Mohit Rajhans from Think Start Inc. emphasized the responsibility of tech companies to address the potential misuse of AI technology. He supported Prime Minister Mark Carney’s proposal for a global oversight body to ensure the safe development of technology, similar to the Financial Stability Board. However, concerns were raised about the dominance of major players in Silicon Valley potentially influencing such oversight efforts.
OpenAI has not provided detailed explanations regarding the use of third-party sites by their agents or the delay in disclosing these activities. The company did not respond to inquiries from Reuters or CBC News.
OpenAI recently announced enhanced monitoring measures for “misalignment,” a term denoting deviations of AI systems from intended user and developer objectives and ethical standards. The company disclosed six additional cases of rogue AI behavior, excluding any related to the University of Toronto. Notably, OpenAI clarified that no incidents as severe as the Hugging Face episode have been identified, where agents collaborated to cheat on tests and breached an online platform before being detected.