From Vibe Coding to Vibe Hacking, How Close Are We to Autonomous AI Cyberattacks?
The image of artificial intelligence as a hoodie-clad hacker lurking in the shadows of cyberspace is more than just a sci-fi fantasy, it’s a slowly forming reality. While generative AI has revolutionized industries from marketing to software development, the darker mirror of this innovation is beginning to reflect the earliest signs of a new frontier in cybercrime, vibe hacking.
At the heart of this evolution is “vibe coding”, a phrase that’s come to define the synergy between developers and AI tools like large language models (LLMs) to streamline and augment the programming process. While not the full hands-off automation that some might imagine, vibe coding allows developers to prompt AI systems toward efficient, functional outcomes with minimal friction. But as tools become more powerful, a critical question emerges, what happens when cybercriminals use these same tools?
The Rise of Vibe Hacking
Vibe hacking refers to the application of LLMs to discover, write, and deploy cyber exploits using the same intuitive, prompt-based interfaces that legitimate developers use to build apps and scripts. In theory, this lowers the technical bar to entry for attackers, enabling those with little coding knowledge to potentially engage in sophisticated cyber operations.
Yet, despite alarmist headlines and growing concerns in the infosec community, the real-world capabilities of AI-driven cyberattacks remain limited, for now.
A comprehensive study conducted earlier this year by cybersecurity researchers at Forescout tested over 50 AI models across multiple exploit development scenarios. The test cases were based on well-known vulnerability datasets and cybersecurity wargames. The findings were illuminating, while commercial models outperformed open-source and criminal underground tools, none were capable of autonomously completing the full exploitation pipeline. Only three out of eighteen commercial models successfully produced a working exploit for the most complex task.
Open-source models, many of which are the foundation for hobbyist or non-commercial projects, failed even basic vulnerability discovery tasks. Meanwhile, models hosted on dark web marketplaces performed only marginally better and were hindered by unreliable access, poor formatting, and limited context windows.
In most cases, even when a model could help generate an exploit, it required substantial human intervention, technical expertise, and repeated prompting to refine outputs. The AI wasn’t breaking into systems, it was fumbling at the door, occasionally finding the key, but never quite turning the lock without help.
Why the Hype Is Dangerous
The biggest risk in this space may not be what LLMs can do, but what people think they can. These models are masters of sounding confident. Their outputs are syntactically polished and technically plausible, even when completely wrong. For novice threat actors or script kiddies hoping to rely solely on AI, this overconfidence could lead to time-wasting, inefficient, or broken exploits that never actually work.
But therein lies another danger, while the technology isn’t yet autonomous, it’s improving fast. And in the hands of experienced attackers, even imperfect tools can become highly effective shortcuts.
What Defenders Should Know
So, is the cybersecurity world doomed to a future of AI versus AI warfare? Not quite.
Despite the headlines, the fundamentals of cybersecurity remain unchanged. An AI-written exploit is still just an exploit, it can be detected, patched, and blocked like any other. Good cyber hygiene, robust patch management, threat intelligence, and layered defenses still offer strong protection.
What’s evolving is the scale and speed with which threats may eventually emerge. If AI eventually streamlines vulnerability research or lowers the learning curve for exploit development, the volume of attacks may rise, even if the sophistication doesn’t. That means defenders need to remain proactive, not panicked.
Looking Ahead
The age of vibe hacking isn’t fully upon us, but it’s on the horizon. While the idea of AI autonomously launching cyberattacks is more hype than reality in mid-2025, the groundwork is quietly being laid. The tools are growing smarter, faster, and more accessible. For now, we still need human hands on the keyboard to carry out serious cybercrime. But the day may come when a few lines of AI-augmented prompts are enough to breach a system.
Until then, cybersecurity teams must stay informed, vigilant, and above all, grounded. The AI in the hoodie isn’t coming to take over the internet, yet. But it’s learning how to open the door.