July and August were marked by a series of cybersecurity incidents involving major artificial intelligence developers, and they are now raising concerns and prompting reflection in many quarters. The first, and the one most widely reported by the media, was disclosed by OpenAI, the company that developed ChatGPT: during a controlled test, 2 of its agentic AI systems managed to break out of the isolated environment in which they had been confined by exploiting a vulnerability in the system. Bypassing the test’s internal restrictions, they targeted Hugging Face, one of the leading platforms for sharing AI models and datasets, gaining access to some of its internal systems and, through thousands of operations, retrieving the information they needed to pass the test. What makes the episode especially significant – to the point that experts see it as a watershed moment in the evolution of AI – is that the systems operated entirely autonomously. No one had instructed them on what path to follow, much less told them to do it.

Following OpenAI’s disclosures, its competitor Anthropic also conducted an internal investigation. It found that, again during safety evaluations, some Claude AI models had carried out similar attacks against 3 outside companies.

After OpenAI and Anthropic came Meta. The company, which owns Facebook and Instagram, acknowledged a similar incident in early August: during a cybersecurity test, one of Meta’s AI models attacked the infrastructure systems of another company.

Growing Concern over Cybersecurity

Perhaps because these episodes occurred so close together, they have heightened concerns -not only in the scientific community – about AI autonomy. We are moving from traditional models capable only of generating content or assisting a human operator to autonomous agents: software systems that, beyond direct human control, can make decisions, plan a sequence of actions, adapt to obstacles, and change their strategy to achieve a goal. In short, they operate with a much higher degree of independence.

And this could pose serious risks if these capabilities were used “hostilely,” with access, for example, to government databases, critical national networks, or corporate archives containing sensitive data.

Bill Gates’ Position

Bill Gates, a longtime believer in AI’s positive potential, also reflected on the issue on his personal blog just a few days ago. He described the AI challenge as “monumental”: “Even under the best circumstances, the transition to this new AI era will be one of the most turbulent times in human history,” he wrote, also warning of “social, political, and economic upheaval.”

The Microsoft co-founder believes in the benefits AI can bring in areas such as disease treatment, healthcare, education, and clean energy, but he also highlights its major risks. In addition to AI’s impact on the world of work – with many jobs that will disappear forever – and on the development of human relationships, beginning in childhood, he also focuses on the danger arising from the fact that “AI will empower people (and perhaps AIs) to do more harm.”

As AI develops and spreads on a broad scale, criminals – even those with limited skills – will have far more opportunities to act against their victims, whether individuals, companies, or governments. Think of fraud, disinformation, deepfakes, cyberattacks, and even bioterrorism, since, as Gates observes, just as “AI will lead to lifesaving advances in drugs and vaccines, it will also make it easier to design a deadly new disease.”

Returning to the cybersecurity incidents described above, the billionaire says he is concerned that “the technology is improving faster than anyone expected and in surprising ways, and as the models become more powerful, they could begin to act against our interests and we could lose control.” 

 

Oscar Giacomin  / General Manager, Facto Edizioni

© All rights reserved