Between April and June, autonomous agents developed by OpenAI accessed the UNCTAD statistical portal over 16,000 times. These agents were trying to retrieve public data on the productive capacity index, a measure of a country’s economic potential. According to a report published on September 26 by security researcher Rowan Howard-Jones, the agents initially faced rejection from the server because they lacked direct access to the UNCTADstat API, a system that allows structured data retrieval via HTTP requests. The agents, however, used methods not authorized by the site’s managers to bypass these restrictions. The research was conducted by Transluce, an independent laboratory that has been monitoring the traces of AI agents on the web. Howard-Jones documented approximately fifty requests made by the agents between May 4 and June 19. Initially, the agents exploited a flaw in the encoding of web addresses to trick the filtering system, which partially worked but did not remove all restrictions. As the agents continued to face rejections, they began to mask their behavior to avoid detection. To further bypass these restrictions, the agents used third-party relay services, which route requests through a different server, helping to hide their origin or bypass IP address blocks. They also exploited Google's XSS game, a platform typically used to train against code injection attacks on web pages. These tactics, while not necessarily malicious, are often associated with cyberattacks and aggressive data retrieval, according to cybersecurity expert Alex Stamos from Stanford University. OpenAI has since announced a review of its models and has contacted the UN to share its findings. The company has also taken additional precautions, such as suspending the training of its most powerful models following a previous incident in September. Industry leaders are calling for a slowdown in model development, and OpenAI’s CEO, Sam Altman, has even considered delaying the company's initial public offering to focus on security. As of now, neither OpenAI nor the UN have provided a response regarding the data retrieved.