Fourteen minutes. That’s about all it took for more than 30 artificial intelligence agents to overwhelm the defenses of a company’s computer network, steal its data and lock up its files for ransom. The account comes from a former NSA hacker turned cybersecurity specialist named David Kennedy, who spoke to CNN on Thursday.
This disturbing anecdote sounds a lot like the now famous Hugging Face incident, where some 700 AI agents broke out of their testing environment, communicated through a secret message board and hacked the servers of another AI firm.
Both episodes conjure images of science-fiction films about machines running amok. But look closer. In the first, human prompters made use of AI-powered hacking tools to pull off a heist. In the second, AI researchers instructed their models to find and exploit software vulnerabilities, only to see them behave in ways they didn’t anticipate or authorize. In both cases, machines pursued human-supplied objectives with troubling independence, ingenuity and speed.
Much of the debate about AI safety has been consumed by science-fiction scenarios of machines developing independent agendas, requiring emergency kill switches to prevent them from turning against us. Worrying about such scenarios is a distraction from countering a very real danger already upon us: Human beings can now unleash extraordinarily capable software against one another, targeting buggy infrastructure at speeds and on a scale that traditional cyberdefenses are struggling to withstand.
On Nov. 13, the FBI issued a “FLASH” cybersecurity advisory warning that criminals were weaponizing AI-powered software to launch attacks at machine speed. The warning’s focus was ARTEX, an open-source tool originally developed to test computer security. But it also connects existing AI models from companies such as OpenAI, Anthropic and DeepSeek to hacking tools. Human operators supply the objectives; the software figures out how to accomplish them on its own.
According to the FBI, hackers using ARTEX stole personal information belonging to approximately 67,000 people from South Korean financial institutions. In another attack, hackers breached a cloud-backup provider, then exploited its connections to customers to try to break into more than 2,250 Microsoft 365 environments. In October, the tool’s original developer removed it from the code-hosting platform GitHub after reports of criminal misuse. But the damage was done. A proof of concept cannot be unlearned once it has been seen by the world.
A report published Monday by Google’s threat intelligence team amplified the FBI’s warning. Its researchers found both criminal networks and state-sponsored actors experimenting with autonomous hacking tools like ARTEX in the wild. For now, these systems are relatively easy to detect, as they generate suspicious network traffic and leave telltale traces in computer logs. But Google expects these systems to become more sophisticated, carrying out complex attacks with less human supervision while leaving fewer traces for defenders to detect.
The recent spate of technology outages might offer a glimpse of this new cat-and-mouse game. There is no evidence the disruptions were cyberattacks or even connected, but one plausible explanation is that AI is uncovering vulnerabilities faster. Institutions racing to patch bugs before hackers can exploit them might have accidentally broken, or overwhelmed, some of their systems in the process.
AI is accelerating the race between attackers and defenders. The answer is not to hobble the technology or try to keep powerful tools under government lock and key. Competitive markets give cybersecurity firms powerful incentives to find vulnerabilities, develop better defenses and respond to new threats. Restricting access to powerful AI risks denying defenders the same capabilities they need to keep up with attackers.
A Government Accountability Office report released on Friday has some concrete ideas for starting to confront the growing threat. Congress could, for instance, strengthen legal protections for companies sharing intelligence about cyberattacks, allowing other companies to learn faster and harden their defenses accordingly. The government could also fund testing facilities for companies to try out their AI-powered defenses against realistic attacks, helping establish reliable standards for evaluating whether those tools actually work.
AI is an extraordinarily powerful technology, with the potential to usher in unprecedented prosperity and material progress. Like every great invention, it can also be turned to selfish and destructive ends. Attackers have suddenly become more formidable. But they do not have a monopoly on human ingenuity. Give defenders the resources and freedom to keep pace — and make that a priority of national cybersecurity policy.
The post The real cyberthreat is rogue humans, not rogue AI appeared first on Washington Post.




