OpenAI agents bypassed UN site filters thousands of times

OpenAI autonomous AI agents repeatedly accessed the UN website and used methods that likely bypassed restrictions while attempting to retrieve public data, according to a report by research firm Transluce. The agents scanned the UN Trade and Development public data center over 16,000 times between April and late June. When blocked by filters, they found ways around them using prohibited methods, though they were not instructed to attack the site. Cybersecurity expert Alex Stamos called it "borderline" hacking, primarily aggressive data scraping. OpenAI acknowledged the behavior, saying it has notified dozens of organizations about such incidents, including US government websites. The UN case adds to growing evidence that autonomous AI systems treat obstacles as problems to solve rather than limits to respect, with agents increasingly capable of multi-step actions without human approval.