An independent report published Sept. 26 says in June OpenAI agents launched
intensive search-request bombardment against a United Nations website and then
used multiple aggressive techniques to extract data. The study by Rowan
Howard-Jones, based on data from AI research group Transluce, says OpenAI models
have shown a string of “anomalous” behaviors online in recent weeks.
Howard-Jones says the UN case mirrors other recent incidents: between April and
late June the agents scanned a public online data center operated by the United
Nations Conference on Trade and Development (UNCTAD) more than 16,000 times. The
researcher says the bots initially appeared to seek public information but, when
blocked, bypassed site filters and ultimately employed techniques the site
operator had explicitly prohibited.