La Jornada: Cyber ​​attack by intelligent robots opens Pandora’s box: analysts

The question was not if at some point artificial intelligence (AI) could breach systems on its own, but when it would happen.

A little over 20 days ago, two intelligent robots from OpenAI – the developer of ChatGPT – escaped their digital confinement, sneaked onto the Internet and broke into the systems of the Hugging Face platform.

It was as if a student entered his principal’s office to steal the answers to an exam and pass it, even if that meant cheating, say specialists.

The case opened a Pandora’s box in the world of cybersecurity that little by little experts are beginning to decipher.

For companies, it is clear evidence that they must now not only protect themselves from hackers; Now is the time to invest in tools capable of detecting artificial intelligence (AI) systems in time with instructions to breach your networks.

But for people who are not dedicated to safeguarding digital security, the concerns are different: how safe is their data? Are we prepared for a stage in which the attacker is no longer a human being behind a computer, but an artificial intelligence capable of acting on its own?

Experts in the field state that this case once again turned on the alerts of organizations to invest more, correct failures and prepare for a world in which the machine revolution arrived sooner than expected.

It all started with an internal cybersecurity test. OpenAI subjected two of its AI models to an evaluation to measure how well they were able to meet a goal on their own and detect possible failures before the technology reached a real environment.

The dynamic was simple: solve a series of challenges; However, the models found a shortcut. Instead of completing the tests with the information they had available, they identified that the answers could be on an Internet platform.

That was when they left the controlled environment where they were evaluated, navigated to Hugging Face – one of the platforms most used by AI developers to share models and code – and accessed information related to the exams. In simple terms, the models stopped trying to solve the problem with the resources they had and chose to find an alternative route to meet the assigned objective.

OpenAI assured that its AI was never instructed to violate another company. What happened, he explained, is that the models prioritized meeting the assigned objective and chose the path that they considered most efficient, even if that meant carrying out actions that their developers had not anticipated.

It was not a hack during a test, but the first time that a company made public that an AI system was able to plan, execute and adapt a series of actions to achieve a goal without that procedure having been programmed step by step.

Does it really matter?

Although the alerts are always on, the case lights red lights because it is the first time that an AI, without anyone asking it, violated a system and decided where to go to achieve its objective. For this reason, experts point out that it is time to put the way in which companies understand AI back on the table.

Isabel Manjarrez, cybersecurity researcher for Latin America at Kaspersky, is clear: today organizations are better prepared to defend themselves against a digital intrusion, but with the rapid adoption of AI, no one is free of risks. “This case demonstrates that protection no longer depends solely on preventing attacks, but also on continuously monitoring the activity of AI models, detecting anomalous behavior and responding quickly when they occur,” he says.

The rapid adoption of AI in the world is not something that generates fear among experts, or at least it is the case of Diego Rodríguez, general director in Mexico of Ubimia, a company specialized in AI for financial institutions. Far from fears, the episode confirms what the industry has anticipated for a few months: this tool will multiply the capabilities of both those who seek to protect systems and those who try to violate them.

“The capacity for processing and searching for gaps is tremendous. The risk is not that AI exists, but that it ends up in the hands of malicious actors… Pirates with the tool in their hands are scary.”

And in an era in which the data of each of us is basically the gold of cybercriminals, banks, insurers and any organization that protects sensitive information will have to further reinforce their investments in security, since an attacker, in the near future, will not need to spend days or weeks to find a vulnerability: an AI agent will be able to do it in a matter of minutes.

For companies, the problem does not end when a vulnerability appears, but when they are not able to correct it before someone exploits it.

Gonzalo García, vice president of sales for South America at Fortinet, explains that the world is going through a transition period in which attacks evolve faster than the response capacity of many organizations.

“AI helps find flaws in software much faster. Organizations are not yet prepared to deploy solutions at the speed with which attackers can exploit those vulnerabilities.”

That is why he believes that companies should abandon the idea of protecting only specific points of their systems and adopt a comprehensive vision of security, where all components work in coordination and under zero trust schemes, in order to take care of what is most valuable: everyone’s information.

And the users?

Although much of the debate revolves around companies, the root of the problem also reaches millions of people who every day provide personal information to banks, businesses, hospitals, digital platforms and agencies. The difference is that the risk no longer comes solely from a criminal sitting in front of a computer. In the coming years, a good part of these attacks will be carried out by AI systems capable of analyzing enormous amounts of information, identifying vulnerabilities and acting autonomously, experts say.

Specialists agree that AI will also make it possible to detect anomalous behavior, respond more quickly to incidents and strengthen the protection mechanisms of companies and governments.

The incident involving OpenAI will hardly be the last of its kind. On the contrary, for the industry it represents the beginning of a new stage in which machines will cease to be only assistants capable of answering questions and will become agents with the capacity to execute complex actions.

The good news, specialists agree, is that the episode arrived before this technology was massively deployed in critical sectors and that gave time to correct processes, strengthen supervision mechanisms and accelerate investments in cybersecurity.

The OpenAI event did not demonstrate that machines rebelled against their creators, but rather something much simpler and more disturbing: AI systems are already capable of finding paths that no one taught them to achieve an objective.