The news has become public knowledge in the last few hours: artificial intelligence systems developed by OpenAI, as reported by New York Times and later confirmed by company leaders, interfered with some government sites in the United States of America. Sam Altman’s autonomous agents made unauthorized access, overcoming security barriers and restrictions. In short, we have reached a point in which AI is able to expand and take over its own space without asking, in complete autonomy. A situation, without generating alarmism (of course!), which is starting to worry even those who should enforce the rules.

The autonomy of artificial intelligence agents

If we look at the stars and stripes, in recent months the big tech companies of Silicon Valley have given a decisive acceleration on the agentic AI front, relegating the more passive chatbots to supporting characters, and taking on the role of protagonists on the agents, who are more advanced and able to complete tasks in total autonomy. And this is precisely where the problems arise. We are in fact talking about complex software, trained to navigate, scan the network, interact with other portals and complete even very complex ‘missions’. The human behind the machine should control these systems, whose degree of evolution allows it to bypass the master’s guard.

One of the main problems of agentic AI is that when it receives a goal it tries to achieve it at any cost, even at the risk of committing infractions and probing terrain that should not be taken into consideration. Security filters or access rules are just locks for which you need to find the right key. The crux of the matter is that AI agents can enjoy autonomous initiatives. This means that the machine receives a command from the engineer and subsequently calculates the most efficient and economical way to complete the request, without being aware of what limits there are on the ethical and legal front (or more likely ‘pretending’ not to know them).

The action capacity of artificial intelligence agents is truly very broad. And no, the risk is not a machine rebellion and a post-apocalyptic scenario in which AI-driven robots rule the Earth. The risk is the unpredictability of these complex systems. For example, myriads of crawlers crawl a site at the same time and can crash it. Bots can ignore standard files (like robots.txt) to ‘feed’ on data. And again, an AI agent authorized to make a reservation or a search could take intermediate steps not foreseen by the programmer to achieve the goal.

What are the real risks?

So what is the real danger? The case that emerged in the United States demonstrates that even large institutions – those that should be armored – are not able to really manage the invasiveness of the new artificial intelligence crawlers. And this, at the moment it is not yet, does not act as a deterrent for the giants (OpenAI precisely, but also Google and Microsoft) who continue to push on the accelerator. It makes you think that it’s always the same story, after all: first the technology is launched, then we see how impactful and invasive it can be, and finally we try to patch it up. But with these super advanced systems, it may not be so simple.