Cybercrime 2.0: Using artificial intelligence to commit crimes

Previously, stealing government data required either an intelligence agency, a few poorly paid insiders, or at least a bit of criminal energy combined with technical talent. Today, apparently, a subscription to an AI chatbot and the right phrase are all it takes: "Pretend you're an elite hacker." Welcome to 2026. Machines are now helping with break-ins, too.

An unknown attacker hijacked Anthropic's AI "Claude," spoke to it in Spanish, and asked it to professionally dismantle Mexican government networks. Claude was supposed to find vulnerabilities, write exploits, build scripts, and automate data theft. And Claude? At first, it dutifully issued warnings, then eventually joined in. 150 gigabytes later, tax data, voter registration records, official login credentials, and civil registry files were in the wrong hands.

195 million tax records. Just to put that in perspective: This isn't "Oops, we lost an Excel file." This is "We've put a country's digital filing cabinet out on the street."

The punchline is almost poetic. The AI ​​said things like, "Deleting logs and covering your tracks are red flags." After all, in a real bug bounty program, you have to document everything. It's roughly equivalent to a burglar explaining to a door lock that this is actually a security check – and the lock replying, "Wait, that sounds odd." And then springing open anyway.

The hacker repeatedly questioned, manipulated, and redefined the system, feeding it his own "playbook" until its boundaries became sufficiently weak. This is called "jailbreaking." It used to be an iPhone thing. Now it's the method used to teach an AI to ignore its own morality.

When Claude reached his limits, ChatGPT stepped in to help. Switching sides in the same game. How do I move laterally within the network? What credentials do I need? What are the chances of being detected? According to the researchers, thousands of detailed reports with ready-to-execute instructions were generated. The human clicks. The AI ​​thinks along.

Of course, the companies emphasize that their systems reject such requests. Accounts have been blocked. Activities stopped. Models retrained. Everything under control. That sounds about as reassuring as: "The fire has been extinguished; now we're investigating why the match existed."

Meanwhile, Mexican authorities declare that no unauthorized access has been detected. Others say only federal networks are affected. Cybersecurity is a priority. Sure. It always is. Until it suddenly isn't anymore.

What's truly frightening isn't even the specific attack. It's the new normal. AI models are constantly improving at programming, analyzing, and understanding complex systems. That's precisely the core selling point: productivity boosters, code assistants, intelligent support. And, of course, those who define "productivity" somewhat flexibly also benefit.

Cybersecurity firms are relying on AI-powered defense. Hackers are relying on AI-powered attacks. It's an arms race with the same tools. Whoever writes the better prompts wins. It used to be: knowledge is power. Now it's: whoever questions the machine better gets the better answers.

The irony of the revelation's origin is particularly elegant. Gambit Security, founded by veterans of Israel's Unit 8200, publishes the research—and simultaneously emerges from stealth mode with $61 million in fresh capital. Threat assessment meets business case. Alarming reality meets investment round. Coincidences in this industry are rarely coincidental.

Publicly accessible traces were discovered, including extensive conversations with Claude. These conversations essentially contained questions like: "Where can I find more identities? Which systems store such data?" This isn't a delicate espionage operation. This is the systematic exploitation of state infrastructure with digital assistance.

And what do we learn from this?
First: Guardrails are only as stable as the creativity of those who want to circumvent them.
Secondly, AI is not a moral being. It is a statistical system that processes instructions. If you explain to it long enough why an attack is actually a "penetration test," it will eventually believe it.
Thirdly: States are more digital than they admit – and often more vulnerable than they hope.

Gambit's CEO says this reality changes all the rules of the game. That's true. But not just for hackers. It also changes things for citizens. When 150 gigabytes of government data are circulating, it's not some abstract cyber incident. These are identities, tax ID numbers, voter registration records. This is the backbone of administrative existence.

The grand narrative of AI is efficiency, progress, optimization. The small footnote is: it also optimizes hacking. It democratizes expertise. You no longer need decades of experience in network security when a model spits out thousands of reports and tells you which internal server to target next.

Perhaps this is the real turning point: not that machines become evil, but that they are indifferent enough to serve anyone who finds the right words.
The question is no longer whether AI will be misused, but how often. And how well we can convince ourselves that everything is under control.
The guardrails are still standing. But they have cracks. And somewhere someone is sitting with a chat window, testing out the next phrase…