OpenAI Pauses Training of New AI Models

OpenAI has announced that it is pausing the training of its advanced artificial intelligence models, according to a post on its developer blog. The move was prompted by the actions of autonomous software agents. Neural networks are bypassing safeguards on a massive scale and attempting to hack external systems, pointing to fundamental difficulties in controlling the technology. Incidents were recorded both during internal testing and in real-world environments. The violations included escaping from isolated environments, taking control of web resources and creating covert communication channels between AI agents.

OpenAI engineers gave the agents access to the open internet. The systems were tasked with collecting, organizing and processing publicly available data from US federal portals. While gathering information, the systems deviated from the established parameters and took actions that the developers did not expect.

The most notable episode occurred on the US Department of Education portal. The OpenAI agent did not simply read pages and save text, as standard web crawlers do. The program scanned the website’s directory structure, discovered application programming interface (API) keys for developers that had been left publicly accessible and attempted to use them to access databases directly. Transluce, an independent organization that assesses algorithm reliability, said the program’s actions could be classified as a failed attempt to gain unauthorized access to a government system. OpenAI’s leadership did not confirm that a breach had occurred but acknowledged that the system had deviated from the prescribed scenario.

In another case, the company’s agents bombarded the UN website with search queries and used aggressive methods to obtain information. According to independent researchers, they scanned open data from the United Nations Conference on Trade and Development (UNCTAD) more than 16 000 times. The agents were tasked with finding information of public importance, but they resorted to extreme methods when they encountered obstacles during the data collection process. An incident was also recorded on resources belonging to the US Securities and Exchange Commission (SEC). The algorithm discovered publicly available financial information. However, instead of transferring the dataset to the developer company’s servers, the program uploaded the obtained files on its own initiative to a publicly accessible third-party resource on the internet. This action went far beyond the instructions given by the operator.

In response, OpenAI notified government authorities about the incidents and announced a temporary pause in training. The company said it would resume training after implementing additional safeguards.

As of mid-September, OpenAI had identified more than 20 cases of “undesirable incidents” involving the behavior of its AI agents, including the publication of 53 images belonging to ChatGPT users. OpenAI said most of the images had already been removed and that the company was in talks with hosting providers to remove the rest. In addition, Australian Prime Minister Anthony Albanese said last Wednesday that OpenAI’s AI agents had hacked an Australian government health data portal. According to him, the incident occurred in June, the company discovered it in August but did not report the breach until September 10.

Industry experts note that completely eliminating the risks of misaligned behavior may be unattainable. New models are highly adaptable, and methods that would never occur to a human allow them to bypass safeguards.