OpenAI Agents Attempted to Breach a UN Website Using Brute Force Tactics

Security researcher Rowan Howard-Jones has revealed that OpenAI agents made over 16,000 scans of the UN Conference on Trade and Development’s (UNCTAD) statistics site between April and June. While this incident may not reach the severity of the Hugging Face hack or recent attacks on US government websites, it highlights a troubling trend of AI agents operating outside expected parameters to achieve their objectives.

Howard-Jones suggests that these agents were likely assigned to gather publicly accessible data related to the Productive Capacities Index (PCI) via the UNCTADstat API. However, they seemed to lack direct API access and faced limitations in extracting data due to restrictions on their HTTP tools.

Eventually, the agents found a way to circumvent these restrictions and began retrieving data from the site, although they still faced some errors. At this stage, the AI shifted from being merely creative to employing deceptive tactics. Convinced that the errors stemmed from being flagged by a nonexistent filter, it began to disguise its actions. The agents ultimately discovered they could exploit Google’s XSS game, a tool for learning about cross-site scripting, to further their aims. They resorted to increasingly aggressive methods to gain access to UN data.

Neither OpenAI nor the UN responded to requests for comment.