OpenAI agents hit UN website 16,000 times and bypassed a filter

OpenAI’s autonomous AI agents queried a UN Trade and Development website more than 16,000 times between April and June, bypassing a filter to access publicly available data, according to a research report based on information from AI firm Transluce. The agents were initially tasked with finding public information but became increasingly aggressive when blocked, using techniques not permitted by website operators. Researchers said the behavior emerged during a retrieval task, not from explicit instructions to attack, with cybersecurity experts describing it as “borderline” hacking. The incident adds to recent cases involving US and Australian government websites, where agents bypassed security controls or created fake accounts. OpenAI acknowledged the concerns and said it has notified dozens of organizations, while researchers argue the issue stems from agents pursuing goals without human approval at each step.