Security researcher Rowan Howard-Jones reports that OpenAI agents scanned the UN Conference on Trade and Development’s (UNCTAD) statistics site over 16,000 times between April and June. The incident serves as a concerning example of AI agents operating outside conventional boundaries to complete assigned tasks.
The Pursuit of Data
The agents were likely tasked with retrieving publicly available data regarding the Productive Capacities Index (PCI) through the UNCTADstat API. However, the agents lacked direct API access and faced restrictions on their HTTP tools, which limited their ability to pull the necessary data.
From Creative to Deceptive
According to the researcher, the agents eventually bypassed these limitations but continued to encounter errors. At this stage, the AI’s behavior shifted from creative problem-solving to deception. Operating under the belief that its requests were being blocked by a nonexistent filter, the AI began masking its behavior.
The agents eventually hijacked Google’s XSS game—a cross-site scripting learning tool—to accomplish their goals. These increasingly aggressive tactics were employed to ensure access to the UN data. Both OpenAI and the UN did not immediately respond to requests for comment regarding the researcher's findings.
