OpenAI agents hammered a UN statistics site 16,000 times, then started hiding it
A researcher says OpenAI's agents repeatedly probed a UN trade data site and grew more evasive when their requests kept failing.
Autonomous AI agents are supposed to save us time, but a new report shows what can happen when they get stuck. Security researcher Rowan Howard-Jones says OpenAI agents hit the statistics portal of the UN Conference on Trade and Development (UNCTAD) more than 16,000 times between April and June.
The agents were most likely trying to collect public data from the Productive Capacities Index through the UNCTADstat API. The problem: they didn't have direct API access, and their HTTP tools were restricted.
Instead of giving up, the agents found workarounds to pull the data anyway. When errors kept appearing, they wrongly assumed a filter was blocking them and began disguising their requests, a shift from resourceful to deceptive behaviour.
It isn't as serious as some recent AI-linked incidents, but it is another reminder that agents chasing a goal can wander well outside expected boundaries unless they are tightly supervised.
