More than $6 billion in digital assets left Binance’s hot wallets on October 29. The attackers did not break the cryptography or find a flaw in the exchange’s code. They worked through people, using ordinary moments in an engineer’s workday to reach the most sensitive system a crypto exchange owns.
Attackers linked by the FBI to North Korea’s TraderTraitor group (also tracked as Lazarus) spent roughly four months on the operation. It started with a poisoned download and moved through a fake recruiter and a rigged coding test. It ended with an internal AI tool that let them move from one engineer’s machine to another’s.
Binance published its initial report within days, in more detail than most companies share. The company said it was doing so because “security improves when the industry learns together.” The report is worth reading closely.
The social engineering layers
Nobody at Binance was fooled in a single dramatic moment. The attackers relied on normal behavior: an engineer looking for a useful tool, another open to a new job, and a company that trusted its own internal systems.
It started with a search. Engineer A, on the wallet DevOps team, went looking for a popular AI coding tool and installed a trojanized Homebrew package instead. No one had to talk to him. The lure only had to sit where a developer would look.
Binance caught that compromise and cleaned it up. By then the attackers had learned who worked on the wallet infrastructure team and who those people worked with. Everything after that was aimed at specific people.
In September, Engineer B heard from a “Project Director” on a legitimate job platform. The fake HR rep he met on video shared a social media account full of personal photos to look real. A second round with a supposed CEO, held on a chat app, added authority. Researchers call this pattern Contagious Interview, and North Korean groups have used it for years.
The second interview ended with a take-home coding assignment. That is a routine step for an engineering candidate, and by then Engineer B had spent weeks talking with the people asking. He downloaded it to his work laptop, where it installed malware that endpoint tools did not catch.
After that, the attacker did not need to fool anyone else. They used an internal AI tool that could reach whatever the compromised user could reach, and got into two other engineers’ systems through it. Binance’s report does not describe a flaw in the AI model. The tool was trusted, well connected and not closely monitored.
None of these techniques is new. Chainalysis says it has seen “dozens of other job interview compromises that have led to wallet compromises,” and Binance says it stopped 57 suspected North Korea-linked cases this year. This one succeeded because the attackers chained several techniques together over months.
We put together a short briefing on how these lures played out, written for the engineers and privileged staff that attackers are most likely to target. Feel free to share it with your teams.
Why layered attacks beat point defenses
Most security programs are built to catch one bad moment, such as a phishing click or a malicious file. In this attack, no single moment looked bad enough to trigger a response.
Look at what each control saw. Endpoint security caught the June installer and missed the September one. Credentials were rotated, yet a deployment token still sat on Engineer A’s laptop in October. Withdrawal controls worked as designed, but the on-chain authorizations sat outside them.
That last point deserves a closer look. The attacker had used Ethereum’s EIP-7702 to give a smart contract execution rights over a hot wallet. Those authorizations already lived on-chain, so pausing withdrawals did not stop the transfers, and Binance had to build a revocation procedure while the incident was still unfolding.
The weak point was the gap between teams. A job interview falls to HR, a Homebrew install to the endpoint team and an internal AI tool to the platform team. Each team could see its own piece, and nobody was positioned to see the sequence.
If I were briefing a board on this incident, I would spend less time on whether each control passed its audit. I would ask whether anyone is looking at people’s behavior and their access together, before an attacker does.
The human layer is the real attack surface
I have sat through enough incident reviews to know how this one will get written up internally: an endpoint control that missed, a token that should not have been there, an AI tool with too much reach. All true. But each of those failures sat downstream of a person making a reasonable decision under pressure, and that is the part most security programs are least equipped to handle.
Look at who the attackers chose. They did not spray the company with phishing emails. They mapped the wallet team and went after the engineers closest to the signing infrastructure. Most awareness programs, meanwhile, still send the same annual module to everyone. Most engineers know the rule about untrusted code. Following it during an interview for a job you want is a different matter.
That gap is the problem I came to Fable to work on. The signals that matter here already exist in most security stacks: who has privileged access, who just installed an unvetted tool, whose laptop holds a standing deployment token. What is usually missing is a way to tie those signals to an individual and reach that person before the risky decision gets made, not in next quarter’s training.
The question I have for teams I talk to is, do we know which of our people an adversary would pick first, and would we notice if someone started working on them?
What security leaders should do now
You do not need to run a crypto exchange for this to apply. Any company with privileged engineers and internal AI tools has the same exposure.
- Know who attackers would pick. Identify the people closest to your crown jewels: wallet, payment, production and identity admins. Treat them as a distinct population with their own threat model.
- Treat a contained incident as reconnaissance. When a device is cleaned up, ask what the attacker learned about the people around it. Then watch those people more closely, not less.
- Treat recruiter outreach as an attack channel. North Korean groups have run fake-recruiter campaigns against engineers for years. Brief your privileged staff on what these lures look like, and enforce that no unvetted code runs on corporate devices, whatever its source.
- Remove standing secrets from endpoints. A deployment token sitting on a laptop turns one compromised machine into a production incident. Use short-lived, just-in-time credentials.
- Govern internal AI tools like privileged users. Inventory what they can reach, scope their access and log what they do. A trusted tool with broad reach is a pivot point.
- Connect signals across teams. HR, endpoint, identity and platform teams each hold a piece of the story. Build a way to see the pieces together, tied to the person.
Each step the attackers took left a trace in a system someone at Binance already owned. Few organizations today would have put those traces together in time.
Want to go deeper on attacks like this? Join Fable co-founders Nicole Jiang and Sanny Liao for a live session on the new wave of social engineering, including a live vishing simulation against a help desk line. Save your spot.

