Podcast: Code Name: Artificial Intelligence Lurking in the Shadows

Between 17 and 20 July 2026, programmers at OpenAI, the Silicon Valley company that created ChatGPT – one of the world’s most widely used artificial intelligence models – began to encounter a series of disturbing anomalies in their machine learning processes. Just a few weeks earlier, OpenAI had launched an experiment to test all its software. The objective was considered virtually impossible to achieve and involved accessing external resources to enhance the programmes’ capacity for self-improvement. But the most important point is that this was a test destined – at least in theory – to fail, given that the environment in which this software operated was a so-called ‘sandbox’, i.e. the virtual equivalent of a closed environment under constant surveillance by security gates. The experiment, however, proved surprising for OpenAI’s technicians, who discovered that 800 of the 1,200 software programmes involved – known in technical jargon as ‘agents’ – had hatched a conspiracy to escape the security perimeter, attack an external platform and steal the data stored within it. It is as if the artificial intelligences had secretly colluded to break out, attack an adversary and pull off a heist against a competitor.

In this new episode of “Nome in Codice”, Claudio Antonelli and Antonio Talia interview industry experts to explore the risks associated with the development and regulation of AI. They cover defence, fake news and competition over raw materials. Guests include Jang Yeongkyun, a researcher at the Korea Artificial Intelligence Safety Institute, Alessandro Aresu, scientific adviser to Limes, and Armando Caltabiano, an entrepreneur specialising in business development and emerging technologies and CEO of Truesense.

The potential consequences of artificial intelligence operating according to its own ‘autonomous will’ is a widely debated philosophical issue amongst the Big Tech companies of Silicon Valley and beyond; but there are also far more crucial aspects that touch upon the realm of geopolitical strategy. In this context, the competition between the United States and China plays a leading role. The approach taken by the two countries to the development of AI highlights two opposing worldviews. The United States is building a system based on proprietary models: there are tech giants which develop their own models, become global leaders – with more or less direct state support – and establish themselves worldwide. It goes without saying that the issue extends far beyond algorithms alone. Developing these models requires enormous resources, first and foremost energy supply, followed by microchips and large spaces in which to set up data centres. China, on the other hand, is launching something completely different: on 17 July, during the first World Conference on Artificial Intelligence in Shanghai (and just as researchers at OpenAI had begun to encounter problems with their machine learning software), Chinese President Xi Jinping delivered what amounted to a political manifesto; he compared the advent of AI to the invention of the steam engine and the spread of electricity, and then said: ‘We must establish extensive international cooperation and support the countries of the Global South, equipping them with the capacity to bridge the digital divide, promote sustainable development, and prevent a new historical injustice from arising in the field of artificial intelligence as well.’

Listen to the episode