📣 The Voice of 200 CISOs: where identity programs break.
All Posts
November 15, 2026
5 min read
Five Hops From a Fake Job Interview to $6.34 Billion Stolen From Crypto Giant Binance
Binance’s $6.34B theft was a chain of identities, each able to reach the next. After the first infection, nothing was exploited.
Boxes in a grid with a coral path hopping through six of them to a selected last box
Table of contents
Overview

What happened

Binance, one of the world’s largest cryptocurrency exchanges, has now published how the largest crypto theft on record happened. It wasn’t one laptop. It was a chain of identities, and every link was access that already existed.

On October 29, 2026, at 20:05 UTC, an attacker began draining Binance’s hot wallets, the online wallets that process customer withdrawals, across ten blockchains. Most of it was gone in about 20 minutes. The total, $6.34 billion, is roughly four times the $1.5 billion Bybit theft of 2025.

  • $6.34B taken from hot wallets
  • ~20 min to drain most of it
  • $2.85B frozen or intercepted
  • 98% of affected tokens replenished

The haul included 18,270 BTC, about 200,000 ETH and billions in stablecoins. Binance is absorbing the loss, so customers aren’t out of pocket. The damage still went further: withdrawals were frozen for days, and on the first day of reopening, net customer outflows hit $3.6 billion, 2.1% of assets under management. The FBI attributes the attack to North Korea’s TraderTraitor group.

How they got in

The attack looks sudden. It wasn’t. Binance’s November 3 report describes a campaign built over months, and co-CEO Richard Teng told CNBC the attackers “spent months building their way in.”

  • June. An engineer on the wallet DevOps team installs a tampered Homebrew package while looking for an AI coding tool. Binance catches it, replaces the device and rotates credentials. By then, the attackers know who sits on the wallet team.
  • September. A “recruiter” approaches a software engineer on a job platform. Two fake interviews later, the second ends with a coding assignment carrying custom malware that Binance’s endpoint tools don’t detect.
  • October 20. From that laptop, the attacker uses an internal AI tool to take over two more engineers’ machines. On one of them sits a token with container-deployment permissions.
  • October 29. The attacker uses the token to deploy a container, and the container reaches the internal wallet signing service. Funds start moving outside the normal withdrawal process.
Attack timeline

Months to build, minutes to drain

Four stages of patience, then about 20 minutes to move $6.34B.

Blog

One more detail shows how deep the access ran. The attacker also delegated control of a hot wallet on Ethereum (EIP-7702), and those permissions kept working after Binance suspended withdrawals. Because the wallets use MPC, no single key could revoke them, so Binance had to build a revocation procedure on the fly.

Nothing was broken into

Now read the chain again, this time for what is missing. After the first infection, there is no exploit.

  1. A fake-interview laptop reaches an internal AI tool.
  2. The AI tool reaches two engineers.
  3. One engineer holds a token.
  4. The token deploys a container.
  5. The container reaches the signing service.

Each hop worked because one identity could reach the next.

Identity reach

Five hops, zero exploits

Each identity could reach the next, so the chain never needed a break-in.

Blog

You will never stop every fake recruiter. Binance says it detected and blocked 57 suspected North Korea-linked attempts this year, and one got through. Compromise will happen. Reach is the part that is yours, and it is the difference between a bad day and $6.34 billion.

How Oak would have stopped it

Oak builds one live, activity-based identity graph of every human, machine and AI agent, from evidence across your systems: who owns each identity, what it can reach, and whether it is actually used. It shows not only who has access, but who should. Lay that graph over this attack, and every hop in the chain becomes a door you can see and close.

  • The AI tool. An internal AI tool that can take over engineers’ machines, wallet-team machines included, is an AI agent with cross-team reach. In the graph it has a named human owner and a clear map of what it can touch. A path from a general software engineer to the wallet team through that tool is a finding, and it could have been closed before October 20. Afterwards, Binance had to rebuild the tool and its access controls.
  • The deployment token. A token that can deploy containers into the wallet environment is a non-human identity with a dangerous capability. Oak shows its owner, what it reaches and whether anyone uses it, so a token that is broader than its job stands out. Revoke it, and the attacker has no way to launch the container.
  • The path to the signing service. For any identity, the graph answers one question: starting here, what is two, three, four hops away? A route from a fake-interview laptop to the signing service is a path Oak surfaces, so it can be removed before anyone walks it.
  • The aftermath of June. When the first engineer was compromised, the exposure was their role and their colleagues, and scoping it meant guesswork. In the graph, scoping becomes a query: every identity those accounts could reach, and every token to retire. Minutes, not weeks.
Identity graph

Cut one edge, close the path

In the graph, a token that reaches the signing service is a finding you can revoke.

Blog

Seeing the door is only the start. Oak closes the loop: each finding arrives with its owner and its activity, and is closed through a ticket, or fixed directly where you grant write access. Every change is explained, approved by your team, reversible and on record.

The malware still lands. But a compromised laptop that can reach nothing that matters has no path to the signing service, and no path to $6.34 billion.

Oak doesn’t stop the malware or the fake recruiter. The wallet-level controls, such as signing policy, on-chain delegation monitoring and revocation under MPC, belong to the signing stack, and Binance has withheld some details. But as Binance describes it, everything between that first infection and the signing service was identity reach. That is what Oak removes.

{{small-cta}}

Where this leaves us

Fifty-seven blocked. One got through.

You can’t win every one of those. Whether the one that gets through reaches the vault is a decision, and it is the whole job.

Find your five hops
Book a demo

Discover related posts