North Korea just pulled off the biggest crypto theft ever: about $6.3B from Binance on October 29.

The way in was a fake job interview. That part isn’t new. This part is, from Binance’s own report:

The attacker leveraged the compromised computer of engineer B and used an internal AI tool to gain control over the Wallet DevOps Engineer A and another Wallet Software Engineer C.

— Binance incident report

Binance’s own AI tool turned one hacked laptop into three.

How It Went Down

  • June: An engineer searched for an AI coding tool and installed a fake version. Binance caught it and replaced his laptop. Too late: the attackers now knew his team.
  • September: Another engineer took a “job interview”. Fake recruiter, fake CEO on video, then a take-home test. He ran it on his work laptop.
  • October 20: From that laptop, the attackers used Binance’s AI tool to take over two wallet engineers’ machines. One held a deploy key.
  • October 29: The deploy key reached the system that signs withdrawals. It signed whatever it was asked. $6.3B left in under four hours.

The Interview Isn’t the Story

North Korea has run fake job interviews for years. Bybit lost $1.5B in 2025 after a developer ran a poisoned project. Binance itself warned about these attacks two days before it got hit.

Someone will always get phished. The real question is how far one phished laptop can get.

The AI Tool Is the Story

At Binance, the answer was: through the AI tool, onto two more machines.

What was the tool? Binance won’t say. Not the product, not what it could access, not how the attackers used it.

You don’t need the details to get the lesson. An AI tool can do whatever its account can do. If that account can reach other people’s machines, so can anyone who takes over the tool. No bug required.

It has happened before. Attackers stole the keys to Salesloft’s Drift AI chatbot and pulled data from hundreds of companies. The s1ngularity malware ran developers’ own Claude Code to hunt for their keys.

What we don't know

Binance says “tool”, not “agent”. Nobody has named it. It could be a coding agent, a chatbot or something homegrown. Anyone who says they know is guessing.

Wrong AI to Blame

Expect “AI-powered hackers” headlines. CZ is already calling for an alliance against AI attacks. There is no evidence the attackers used AI at all.

The AI problem here is Binance’s own tool. It’s the AI scapegoat again: point at scary outside AI, not your own setup. Binance’s fix includes more AI, too: “AI-based behavioural anomaly detection”.

Check Your Own Setup

If you run AI agents at work:

  • What can it reach? Just your laptop, or other machines, deploy keys and prod?
  • Whose account does it use? Yours, or a shared one with more access?
  • How fast can you shut it off? Binance needed almost four hours to kill one bad permission.
  • Where do take-home tests run? Not on your work laptop.

Someone will always get phished. Whether it costs one laptop or $6.3B depends on what your tools can reach.

Related Posts