Skip to main content
← Back to Articles

Case File: The Binance Hot-Wallet Theft

11 November 2026 · Ben Visser · 5 min read

At a glance - TL;DR

On 29 October 2026, about USD 6.34 billion left Binance's hot wallets without any cryptography being broken. Binance's own post-mortem traces the path back through a counterfeit AI coding tool in June, a fake job interview in September and an internal AI tool used to take over engineers' machines in October. This first Case File reads the theft as a chain of reasonable yeses, keeps what is confirmed apart from what is attributed or still unknown, and sets out the ordinary controls that give someone the job of saying no.

Key takeaways

  • Binance's cryptography held; what gave way was everything around the keys: the machines near them, the people with access to those machines and, as far as anyone has said, a signer with no rule about what a hot wallet should never sign.
  • AI played three confirmed roles: a lure (a counterfeit AI coding tool), and both the tool and the target (Binance's own internal AI tool), though how that internal tool was abused is still unexplained.
  • Each step was a reasonable yes where it was said, which is why nobody was standing at that moment with the job of saying no.
  • An internal AI tool's permissions deserve the same treatment as an admin account: granted narrowly, reviewed, and removed when no longer needed.
  • After an incident, ask what the attacker now knows about you, not only whether the machine is clean.

At 20

UTC on 29 October, Binance's wallet signing service started moving money. It had been sent fifty signing transactions from a container inside the network, deployed with stolen credentials, and it did what it was built to do. By one analytics firm's count, most of about USD 6.34 billion had left the exchange's hot wallets within twenty minutes. Nobody broke the cryptography. Nobody needed to.

The attackers had held the machines of two Binance wallet engineers since 20 October. The way to those machines started four months earlier, with somebody trying to install an AI coding tool.

Binance caught that one.

How the chain was built

Read Binance's post-mortem, published on 3 November, in order and the theft stops looking like one event. It is a chain, and each link handed the attackers what the next one needed. Where Binance confirms a link, I say so. Twice I have had to join the dots myself, and I say that too.

June: a foothold that was cleaned up. An engineer on Binance's Wallet DevOps team went looking for a popular AI coding tool and downloaded a tampered Homebrew installation instead. Binance spotted the compromise at the time, replaced the device and rotated the credentials. By the standards of most security teams, that is a good day. But the post-mortem adds one more line: the engineer's role had been exposed to the attacker. The machine was gone. What the attackers had learnt was not.

The first join. Binance does not say what the attackers did with that knowledge. My reading: once you know who works on the wallets, you know where to aim next. That is my inference, not Binance's finding, and I hold it loosely.

September: a job offer. A second engineer was approached about a job. A fake HR representative appeared on a live video call, and a fake chief executive followed. Then came a coding assignment, and its repository carried malware onto a work computer. A job interview is built for trust: you want to be quick, helpful and impressive, and nobody opens a take-home exercise expecting it to bite. The attackers now held a machine inside Binance. The report does not say how this engineer was chosen.

20 October: from a machine to the people near the money. With the machines they held and an internal AI tool, the attackers took control of the devices of two wallet engineers. Binance has not said how, and that gap sits at the centre of the chain: the step from someone inside the company to the people closest to the keys.

Nine quiet days. For the next nine days the post-mortem reports nothing. Then the attackers had credentials that worked. Binance says they were stolen; that they came from those two engineers' devices is my second join.

29 October: the signer does its job. From 17

to 19
UTC the attackers were inside. They deployed a container that could reach the internal wallet signing service, sent the fifty signing transactions and set malicious delegations. The signer did the rest. Binance gives the last transfer as 23
.

After the pause. Two analytics firms describe a second wave on Avalanche and Polygon, around 23

, after Binance had paused withdrawals. Binance's post-mortem does not mention it. The last step on record is Binance's own: at 00
UTC on 30 October, it revoked the delegations.

By 3 November Binance reported about USD 2.85 billion frozen or intercepted, and said it had set aside more than the full loss. The FBI attributed the theft to North Korea, which Binance relays in its report, and coverage of the FBI's alert names the group as TraderTraitor. North Korea denies it.

This is the first of our Case Files: a real incident, read for what it teaches rather than who to blame. Every one of them was somebody's worst week. Binance kept its engineers anonymous, and so do I.

A chain of reasonable yeses

Walk back along that chain and every link was let through by an ordinary yes: the install, the coding test, and a signing service that received fifty requests from inside its own network and, as far as the post-mortem tells it, signed them.

Each yes was reasonable where it was said. Call it a chain of reasonable yeses. No single link looks like the moment it went wrong, which is exactly why nobody was standing there with the job of saying no.

The ledgers held. Binance's hot wallets used multi-party computation, which spreads signing across several parties so that no single machine holds the whole key, and nothing in the reporting suggests it was broken. What gave way was everything around the keys: the machines near them, the people with access to those machines, and a signer that, as far as anyone has said, had no rule about what a hot wallet should never sign.

None of this is meant to mark Binance's homework. Most organisations would have said every one of these yeses, and I am not certain we would have said no to all of them.

Lure, tool and target

AI turns up in this story three times, in three roles, and Binance's post-mortem confirms all three.

The lure. The first step needed no AI from the attacker, only our appetite for it.

The tool, and the target. The AI tool used on 20 October was Binance's own. The post-mortem calls it "compromised internal AI tooling" and says it has been rebuilt, along with its access controls. No model vulnerability is named, and commentators call the missing how the biggest gap in the report.

What AI did not do matters as much. Nothing suggests an AI acted on its own: people directed this. No source says the faces or voices in those interviews were generated, so I am not calling them deepfakes. And no source credits AI with catching the theft.

What is still unknown

Binance published within five days, which deserves credit. Four questions remain open.

The biggest is how the internal AI tool was abused, and what it could do on engineers' machines. If it held broad permissions of its own, the lesson is about how much power organisations hand their AI tools. If the attackers simply drove it with stolen access, the lesson is about access.

The rest: whether the attackers used AI to research their targets or to run the interviews; the evidence behind the TraderTraitor name, which a fuller FBI release or an indictment would show; and the final amounts stolen and recovered, which wait on Binance's final report.

What compliance would have changed

Most people reading this do not run hot wallets, but almost every organisation has its own version of each link in that chain. The controls that would have mattered are not exotic. They are the ordinary ones, done for real.

Give people a fast, verified route to the tools they want. Install software only from verified sources (ISO 27001 Annex A 8.19), AI tools included, and make the approved route quick enough that searching the internet yourself is never the easier option.

Keep the power away from the places where code gets tried out. Separate development from production (8.31), so that a coding exercise on a work computer, or a container deployed with stolen credentials, does not sit one step from what holds the money.

Give the signer a rule. Your version might be a payment approval, a deployment pipeline or an AI agent with write access. Each needs something that refuses what should never happen, whoever is asking.

Treat an internal AI tool's permissions like an admin account (8.2): granted narrowly, reviewed, and taken away when no longer needed.

Talk about the recruiter. Awareness (6.3) that works is a conversation, not a slide: if someone offers you a job and a coding test, do it on a machine that is not ours. Looking at other jobs is normal, so make it easy to mention.

After an incident, ask what they learnt. Learning from incidents (5.27) means asking what the attacker now knows about you, not only whether the machine is clean.

This is the part of compliance we care most about at Askara Solutions. A policy that says no is easy to write. The work is making sure that no is something a person or a system actually says, on an ordinary Tuesday, to a request that looks perfectly reasonable.

I wrote in April that as the technical layer hardens, the human layer is what is left. Binance's ledgers held. Somewhere in your organisation, a reasonable yes is waiting to be said. Who has the job of saying no, and what does it cost them to say it?