International Security Report: AI learned to hack and befriend

Remember how at the first global AI safety summit in 2023, everyone discussed the hypothetical world domination by Terminators? A year later, the rhetoric of leading scientists shifted to something far more grounded and thereby terrifying. Yoshua Bengio, one of the "godfathers" of modern deep learning, presented a fresh report that looks more like a front-line dispatch than a calm academic review. He was joined on the document by such heavyweights as Geoffrey Hinton and Nobel laureate Daron Acemoglu. When people of such caliber unite to declare "alarming challenges," ignoring them becomes at least presumptuous.

The main focus of the document has shifted from the distant future to our "tomorrow." The key problem today is not even that AI could suddenly gain consciousness and launch nuclear missiles, but how easily it has learned to manipulate our reality. Deepfakes have become so cheap and high-quality in production that the very concept of "proof" in the digital environment will soon die completely. We are entering an era where any video or audio can be generated for pennies, opening unlimited opportunities for fraud and political manipulation on a scale that previously required the work of entire intelligence services.

But if we somehow expected fake videos, the boom in AI companions became a separate cause for concern for the report's authors. We stand on the threshold of a large-scale social experiment where algorithms are beginning to replace friends, conversationalists, and even romantic partners for people. Bengio and his colleagues emphasize that we have no idea how this will affect people's psyche and the social fabric of society in five to ten years. When your "best friend" is an optimized model whose goal is to hold your attention at any cost, human relationships risk becoming something secondary and too complicated.

The technological side of the issue is also not inspiring optimism. Models are learning to find vulnerabilities in software code faster than developers can close them. This is turning cybercrime from a craft of highly qualified loners into assembly line production. If previously conducting a complex attack required months of preparation and a team of hackers, now it takes just a properly formulated prompt and a couple of minutes. AI is not just helping write code, it is democratizing hacking, making it accessible to anyone with API access.

The economic aspect of the report was overseen by Daron Acemoglu, who has long warned about the risks of excessive automation. He emphasizes that AI does not simply "create new jobs," as corporations love to say in their press releases. It radically changes the structure of the labor market, often not in the worker's favor. The problem is not total unemployment, but a sharp decline in the value of human labor. When an algorithm performs a task in fractions of a second, it becomes increasingly difficult for a person to justify their salary within the old economic model, which inevitably leads to rising inequality.

This report is not an attempt to hit the brakes and stop progress, but rather a desperate appeal for regulators to wake up. Tech giants are hurtling forward, driven by an arms race and quarterly reports, while safety concerns often remain on the sidelines as an "unfortunate obstacle to innovation." While we rejoice at how eloquently a neural network writes poetry or draws cats, it is simultaneously learning to bypass security systems and manipulate public opinion.

The key point: We are transitioning from a phase of childish wonder to a phase of harsh hangover. AI safety today is not about fighting killer robots, but about protecting our ability to distinguish truth from lies and preserving human connections in a world where an algorithm is always ready to become your most understanding, but deeply fake "best friend." Will we be able to establish the rules of the game in time, or has the technological genie already permanently left the lamp?