International AI Safety Report: AI Is Learning to Hack — and to Befriend Us

Remember how, at the first global AI Safety Summit in 2023, everyone was talking about a hypothetical Terminator-style takeover? A year later, leading scientists have adopted far more down-to-earth rhetoric, and that makes it scarier. Yoshua Bengio, one of the "godfathers" of modern deep learning, has presented a new report that reads more like a dispatch from the front line than a calm academic review. Heavyweights such as Geoffrey Hinton and Nobel laureate Daron Acemoglu worked on the document with him. When people of this caliber join forces to warn of "daunting challenges," ignoring them is presumptuous at best.

The document's focus has moved from the distant future to our "tomorrow." The main problem today isn't even that AI might suddenly become conscious and launch nuclear missiles. It's how easily AI has learned to manipulate our reality. Deepfakes have become so cheap and so convincing to produce that the very idea of "proof" in the digital world will soon be dead for good. We are entering an era in which any video or audio clip can be generated for pennies. That opens up limitless opportunities for fraud and political manipulation at a scale that used to require the work of entire intelligence agencies.

Fake videos were something we more or less saw coming. The boom in AI companions, however, gave the report's authors a separate reason for concern. We are on the verge of a massive social experiment in which algorithms start replacing people's friends, conversation partners, and even romantic partners. Bengio and his colleagues stress that we have no idea how this will affect people's mental health and the social fabric five to ten years from now. When your "best friend" is a model optimized to hold your attention at any cost, human relationships risk becoming secondary and simply too complicated.

The technical side of the picture doesn't inspire optimism either. Models are learning to find vulnerabilities in code faster than developers can patch them. That turns cybercrime from a craft practiced by highly skilled loners into an assembly line. A sophisticated attack used to take months of preparation and a team of hackers. Now a well-crafted prompt and a couple of minutes are enough. AI doesn't just help write code; it democratizes hacking, putting it within reach of anyone with API access.

Daron Acemoglu, who has long warned about the risks of excessive automation, oversaw the report's economic section. He stresses that AI doesn't simply "create new jobs," as corporate press releases like to claim. It radically reshapes the labor market, often to the worker's disadvantage. The problem isn't mass unemployment but a sharp drop in the value of human labor. When an algorithm completes a task in a fraction of a second, people find it ever harder to justify their salaries under the old economic model, and that inevitably drives inequality higher.

The report isn't an attempt to slam on the brakes and halt progress. It's closer to a desperate call for regulators to wake up. Tech giants are racing ahead, driven by an arms race and quarterly earnings, while safety often gets pushed to the sidelines as an "annoying obstacle to innovation." While we marvel at how smoothly a neural network writes poems or draws cats, it is also learning to bypass security systems and manipulate public opinion.

The bottom line: we are moving from a phase of childlike excitement into a brutal hangover. AI safety today isn't about fighting killer robots. It's about protecting our ability to tell truth from lies and to preserve human connections in a world where an algorithm is always ready to become your most understanding, yet deeply fake, "best friend." Will we manage to set the rules of the game in time, or has the tech genie already left the bottle for good?