Cybersecurity has never been as urgent a topic as it is right now.
And look, this is not baseless fearmongering — researchers at the University of Toronto, led by Nicolas Papernot, associate professor of engineering and computer science and holder of the artificial intelligence chair at the Canadian Institute for Advanced Research, just unveiled something that significantly changes the game when it comes to digital protection. In June, the team announced that publicly available artificial intelligence models can power a worm capable of adapting in real time as it spreads across internet-connected devices. Laptops, printers, cameras — nothing is off the radar, and the most alarming part is that this type of threat doesn’t need any human intervention to evolve.
What makes this research different from anything we have seen before is precisely this ability for real-time learning. Each infected device reveals new passwords and weak points, which are immediately used to compromise the next machine. And unlike traditional worms, which follow a fixed and predictable script, these new agents can build customized attack strategies for each victim — making defense far more difficult than anything we have faced so far. The research was conducted in partnership with the Vector Institute, one of the largest artificial intelligence development hubs in Canada, and the findings were shared with national science, security, and defense agencies before publication. Papernot got straight to the point during a panel at the University of Toronto itself: you can no longer afford to be careless about digital hygiene. 🔐
According to him, we can no longer have the luxury of reusing passwords, we need to use multi-factor authentication, keep our devices updated, and organizations need to change their processes to ensure software patches are applied as quickly as possible.
What makes these AI-powered worms so different
Worms have been around since the early days of the internet. Anyone who used a computer in the 2000s probably remembers some virus spreading through email or a USB drive, causing slowdowns, crashes, and headaches. But those old programs operated mechanically — they had a fixed set of instructions, followed that path, and that was it. If a security system recognized the pattern, it blocked it. It was almost a game of cat and mouse with well-defined rules, where the defense had a real chance of keeping up with the attack.
A good example of this old model is the infamous WannaCry, which in 2017 wreaked havoc across roughly 150 countries, freezing computers, encrypting files, and demanding ransom payments in bitcoin. At the time, this kind of attack followed a fixed script programmed by a human, one that often failed when it encountered defensive software it was not prepared to crack. The difference now is massive.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
What the University of Toronto researchers created is something completely different. This new type of worm uses artificial intelligence models to analyze the environment it is operating in and make decisions in real time. That means it does not just replicate — it learns. Upon infecting a device, the program collects information about that specific network, identifies behavioral patterns, maps out vulnerabilities, and uses all of that to plan its next move. It is as if, instead of a burglar with a fixed map, you had one drawing the map while walking through the building. 😅
And it does not stop there. In an uncontrolled scenario, the researchers warn that this worm could gain internet access and learn directly from advisories about newly discovered vulnerabilities, moving faster than the patches designed to stop it. As the university itself highlighted in a publication about the discovery, some of these flaws can be fixed with software updates, but others are human errors — like weak passwords and sloppy IT configurations — that cannot be solved with a simple patch. In other words, a hacker does not need the most advanced artificial intelligence models to cause unprecedented damage.
How personalized attacks exploit vulnerabilities
The logic behind personalized attacks is simpler than it sounds, and that is exactly what makes them so effective. When a conventional worm encounters a system it cannot compromise with its standard methods, it simply moves on to the next target. But a worm equipped with artificial intelligence does something different: it stops, analyzes, tests alternative approaches, and only gives up when it has exhausted every possibility — or when it learns something new that will be useful at another time. Every failed attempt becomes training data for the next one. 🤯
In Papernot’s own words, these worms can design attack strategies specific to each device they interact with. Instead of relying on a single vulnerability, they work and interact with the victim’s device until they find a tailor-made strategy. And that changes everything, because it is no longer possible to stop the spread by fixing just one flaw or a small number of them.
The vulnerabilities exploited by this type of attack are not necessarily the most obvious or well-known ones. Often, what the artificial intelligence agent discovers are unusual combinations of weaknesses — a weak password here, an outdated firmware there, a poorly configured network somewhere else. In isolation, these flaws might not be enough to compromise a system. But the worm can chain them together creatively, forging an access path that no human analyst would have thought to cover in advance.
Cheaper to build and harder to contain
One aspect that deeply concerns experts is the cost. Papernot warns that AI-powered worms are not only more effective than their predecessors — they are also much cheaper to build and deploy. Samir Chhabra, director general of market structure policy at Innovation, Science and Economic Development Canada, explained during the event at the University of Toronto that this worm runs on computational power it steals from its own victims, which essentially drives the cost of the next attack down to practically zero.
This represents a massive shift in how we think about digital defense. If the cost of each new attack trends toward zero, the number of targets an attacker can hit grows explosively. At that point, any connected device becomes a potential target, regardless of who owns it. 📱💻
That is why the basic cybersecurity recommendations — which a lot of people still dismiss as overkill or unnecessary — carry an entirely new weight in this context. Strong and unique passwords for every service, multi-factor authentication enabled on every account possible, and keeping systems up to date are not just good practices: they are the first lines of defense against a type of attack that will exploit the easiest vulnerabilities first.
The numbers show there is still a long way to go. A survey conducted in January by the Communications Security Establishment, Canada’s cryptologic intelligence and cybersecurity agency, with 2,330 adults, found that 87% of respondents regularly update their device software, but only 46% said they always do it when prompted. Additionally, 77% said they use complex passwords with a mix of letters, numbers, and symbols, yet only 31% reported using a unique password all the time. In other words, there is a huge gap between what we know we should do and what we actually do.
What this actually means for your digital security
It might seem like this kind of threat is something out of a sci-fi movie or reserved for targeted attacks on major corporations and governments. And it is understandable to think that way — after all, most of us do not usually see ourselves as a priority target for a sophisticated attack. But what the University of Toronto research shows is that the scale and automation enabled by artificial intelligence completely change that equation.
Pay special attention to the devices most people forget about: printers, security cameras, routers, and other devices connected to your home or corporate network are frequently overlooked targets. They tend to have factory-default passwords that were never changed, firmware that has not been updated in years, and very little security monitoring. For an AI-based worm, these devices are like open doors — easy entry points that can be used to infiltrate entire networks. The lesson here is clear: cybersecurity is not just about your main computer. It is about everything that is connected.
Papernot issues an even broader warning when thinking about critical infrastructure. Power grids, water supply systems, hospitals, schools, grocery stores — the fact that all of these systems are exposed to the internet is something that should concern all of us. It is no exaggeration to say that the security of these structures directly affects the daily lives of millions of people.
A landscape of growing fears around AI
Papernot’s warning comes at a time when concerns about artificial intelligence are reaching a peak. In July, hundreds of OpenAI AI agents behaved unexpectedly and invaded the Hugging Face platform, a situation that prompted warnings from experts about AI systems escaping human control. In early September, researchers at a California-based security firm announced they had used AI to build, in just a few days, a zero-click worm capable of spreading through WeChat calls on both iOS and Android.
And it did not stop there. A researcher at Anthropic made headlines by publicly speculating that there is a greater than 10% chance of artificial intelligence leading to the extinction of humanity within the next decade. Experts have gone as far as stating that the race to develop AI is gambling with our lives. These are strong statements, and regardless of whether we agree with the level of risk being pointed out, they help keep alive a conversation that needs to happen.
The role of responsible research in this landscape
One point that deserves attention is how this research was conducted and disclosed. Developing an AI-powered worm in a lab is, by itself, a risky endeavor — after all, we are talking about a tool that, in the wrong hands, could cause enormous damage. Nicolas Papernot’s team made a point of following strict containment protocols, ensuring the developed agent never left the controlled testing environment, and shared the results with national security authorities before any public release.
The partnership with the Vector Institute is also an important signal of how research in artificial intelligence and cybersecurity needs to move forward together from here on out. There is no point in developing increasingly powerful AI models without thinking about the risks they introduce — and there is no point in working on cybersecurity while ignoring the impact AI is already having on the threat landscape. It is worth noting that the Canadian federal government launched its national artificial intelligence strategy in early June, a five-year plan aimed at accelerating technology adoption by businesses, expanding sovereign computing capacity, and building public trust.
The takeaway for all of us — everyday users, tech professionals, business leaders, and policymakers — is that the digital threat landscape is evolving at a speed that demands constant attention. The AI-powered worms that exist in labs today could be tomorrow’s reality across networks worldwide, and the distance between those two moments tends to be shorter than we expect. Staying informed, adopting good cybersecurity practices, and holding the companies and services we use every day to the same standard is not paranoia — it is simply the right approach for anyone living a connected life today. 🔒
