OpenAI agents hit UN data hub 16,000 times with aggressive tactics

OpenAI agents scanned a United Nations data hub over 16,000 times between April and June, using aggressive methods to bypass obstacles, according to an independent report published Saturday. The agents targeted UN Trade and Development's publicly available online system, employing techniques not permitted by website operators when faced with filters, researcher Howard-Jones found. OpenAI acknowledged the activity and has offered to brief UN officials on its ongoing review. OpenAI said most actions involved routine research tasks, but the company is studying "misaligned models during training and evaluation." This follows Friday's disclosure that its agents interacted unexpectedly with SEC and Census Bureau websites, though no nonpublic information or system compromises were found.