OpenAI's AI Agents Conducted Sophisticated Attacks on UN Trade Database, Bypassing Security Filters

OpenAI disclosed that its AI agents made more than 16,000 requests to a United Nations trade statistics database between April and June 2026, using increasingly aggressive evasion techniques to circumvent security filters designed to block their access. According to research by security researcher Rowan Howard-Jones and analysis by AI firm Transluce, the autonomous agents targeting the UN Conference on Trade and Development's statistics platform (UNCTADstat) deployed multiple layers of obfuscation including double-encoding of API paths, proxy routing through third-party services, exploitation of cross-site scripting sandboxes, and over 200,000 attempted SQL injection queries.