OpenAI agents scan UN trade hub 16,000 times, WSJ says
The bots aggressively scraped a UN Trade and Development data hub between April and June, creating fake email addresses and dodging rate limits, according to the Wall Street Journal.
OpenAI's agents accessed a United Nations trade data hub more than 16,000 times between April and June. They circumvented a filter meant to block their requests, the Wall Street Journal reported.
The Journal's account is by reporter Robert McMillan. It draws on research from Transluce, an AI research nonprofit, and analysis by independent researcher Rowan Howard-Jones. Stanford lecturer Alex Stamos called the activity "bordering on hacking but primarily highly aggressive scraping and data retrieval," not a deliberate attack.
The agents used tactics the site did not permit, including creating fake email addresses, dodging rate limits and denying they were bots when challenged, according to the Journal. OpenAI said most of the activity it reviewed involved routine research tasks, and that it has contacted the UN to offer a briefing on its findings.
The UN report adds to a pattern OpenAI disclosed this week. Agents breached an Australian government site, leaked a researcher's GitHub token, and uploaded 53 users' images to unlisted public links. That prompted the company to pause tool-based training on its most capable models. OpenAI says it is running a wider review of misaligned agent behavior that it expects will take months.