Repost

Today’s biggest cybersecurity laugh:

A hacker used an open-source AI agent to break into several major Korean banks and steal tens of thousands of high-net-worth customer records.

Then, because his own server had an open public directory, he exposed all of his chat logs with the AI, his configs, even the job-application résumé he had the AI write for him…

Security firm CrowdStrike followed the trail and turned his whole operation inside out

This guy was actually pretty cutting-edge:
He used the open-source penetration framework ARTEX, running a “multi-agent” setup — one agent planned the attack paths, another executed by calling the tools.
The underlying models are even wilder: mainly DeepSeek, with GLM and Grok chained in along the way.

True “cyber assembly-line work”: the AI does the hacking for him, and he just stands by as the foreman.

But here comes the best part —

The server he rented in Hong Kong had no access restrictions at all, so it became an Open Directory. CrowdStrike walked in like it was strolling through its own backyard and saw every single prompt he’d typed in the terminal:

  1. Asking the AI to help him find buyers on Telegram to fence the stolen data
  2. Asking the AI to polish this “attack haul” and write it into his [cybersecurity researcher job résumé]

Dude, you used AI to hack into banks, and you didn’t even lock down the most basic permissions on your own server?

As ridiculous as it all is, the damage was very real:

Shinhan Bank had about 25,000 personal records leaked (including annual income and loan limits)
KB Kookmin Bank and Hana Bank were hit one after another too

It used to take a hacker weeks to probe a few hundred obscure business endpoints. Now you hand it to an AI agent, and it tirelessly breaks down the tasks, tries things, hits errors and retries, all on its own.

Defenders think they’re facing an ordinary script kiddie, when what’s actually on the other side is a whole automated cluster of large models running 24 hours a day.

This is a big wake-up call for every developer and security person:

  1. Session logs are extremely sensitive data too. Your prompts, AI memory and execution logs could turn into your “confession” or a source of leaks at any moment.

  2. Once script kiddies go all-in on agents, the most dangerous spots are often the fringe business entry points that nobody has paid attention to for ages, yet can still reach sensitive data.

The most advanced AI framework paired with the most amateur basic ops: easily the most absurd security spectacle of 2026.