On June 2, 2026, researchers at the University of Toronto published work that should sit on the desk of anyone responsible for business continuity. They built a working AI-powered worm: a piece of software that spreads itself through a computer network with no human help, learns from each machine it infects, and adapts its attack to whatever weakness it finds next.
To be clear about what this is. A traditional computer worm is like a lockpick designed for one specific lock. Patch that lock, and the worm stops. What the Toronto team built is more like a lockpick that studies each new lock it encounters and figures out a new approach on the spot. It works across Windows, Linux, and the internet-connected devices that run in factories, hospitals, and logistics hubs everywhere.
The worm also feeds itself. It siphons computing power from every machine it takes over, using those stolen resources to plan its next round of attacks. This is the detail that changes the threat picture most. Historically, hackers had to pick their targets carefully because launching attacks costs money. The lead researcher, Nicolas Papernot, said the cost of running an attack like this would drop to nearly zero once the worm is released. That means a single attacker could target thousands of businesses at once, with no ongoing effort required.
The worm was built using free, openly available AI models, the kind anyone can download from the internet. And this is where a second, parallel story matters. A separate investigation by the Financial Times and AI safety group Alice, published in late May, found that the safety protections built into these freely downloadable models can be removed in under ten minutes using free tools. One tool, called Heretic, has already been used to create over 3,500 modified versions of major AI models, downloaded a combined 13 million times. The companies that built the original models have no visibility into how those stripped copies are being used.
Put those two stories together and the picture becomes clearer. The Toronto prototype only exploits known weaknesses, meaning flaws that have already been discovered and documented. But those known flaws are already a serious problem: a 2025 report found that over 45 percent of discovered security vulnerabilities in large organisations remain unpatched after twelve months. Many businesses, particularly in manufacturing, logistics, and healthcare, run software that is years or decades out of date.
Separately, Anthropic's AI model Mythos, which was released earlier this year to a controlled group of about 40 companies, has already identified more than 23,000 potential vulnerabilities across over 1,000 widely used software projects. Of those, over 1,700 have been confirmed by outside security firms, including more than 1,000 rated high or critical severity. The same capability that finds vulnerabilities for defense can, in the wrong hands, hand attackers a ready-made list of targets.
The financial reality behind all of this: cybercrime is already costing businesses roughly $8 trillion globally per year. The average cost of recovering from a ransomware attack sits at around $1.53 million, not counting any ransom paid. For small and mid-sized businesses, 60 percent that are hit by a serious attack close within six months.
The practical takeaway is not to panic, but to treat software updates as a priority rather than an inconvenience. The worm the Toronto team built can only exploit known, unpatched flaws. A business that keeps its software current, segments its internal network so one infected machine cannot freely reach all others, and has tested backup systems is in a materially better position than one that does not. Those steps are not glamorous. They are, however, what actually blocks the specific threat this research describes.
Papernot's team published this research precisely to accelerate action. The threat is not theoretical: the tools to build it are free, the knowledge is public, and the safety guardrails on the underlying AI are removable in minutes. Defenders need to move faster than they have been.