What happened

On 29 October 2026, attackers moved about $6.34 billion out of Binance’s hot wallets across 10–11 blockchains — the largest crypto theft on record.

Binance says no private key was stolen. The attackers spent five months getting into the laptops of the engineers who run the wallets, then used Binance’s own signing service to approve the transfers. About $2.85 billion was frozen; Binance absorbed the rest and says users lost nothing.

Metric Value
Impacted $6.34B (Elliptic estimates $8.43B)
Frozen $2.85B, about 45%
Main drain ~20 minutes, 20:05–20:24 UTC
User losses $0
Attribution North Korea's TraderTraitor group (FBI)

How they did it

The attack chain

Six stages, five months, one trusted tool.

01 / 06

TL;DR A tampered AI tool leaks the wallet team.

Swipe the illustration to see each detail →

Stage 01 · Recon

Engineer A installs a tampered AI coding tool. Binance cleans the device, but the attacker now knows the wallet team.

Stage 02 · Lure

A fake recruiter runs two interview rounds and sends a take-home assignment. Its setup code runs on Engineer B’s work laptop.

Stage 03 · Foothold

The malware replaces pyenv, a trusted developer tool, with a same-name copy that opens a remote shell.

Stage 04 · Pivot

An internal AI tool takes the attacker to two more engineers. Engineer A has a container deployment token on his laptop.

Stage 05 · Breach

A malicious container reaches the internal signing service and triggers transfers outside the normal withdrawal path.

Stage 06 · Drain

Ten chains are emptied in about 20 minutes. Ethereum’s EIP-7702 delegation keeps funds moving after withdrawals are paused.

Five months · six steps · no private key stolen

Source: Binance initial findings, 3 Nov 2026.

Attack path

flowchart LR A["Poisoned install
AI coding tool"] --> B["Lure
fake job interview"] B --> C["Malicious repo
take-home"] C --> D["Foothold
pyenv backdoor"] D --> E["Pivot
internal AI tool"] E --> F["Deploy token
signing service"]

Swipe to follow the path →

The breach

  • Malicious deployment. The token let the attackers deploy a malicious container next to Binance’s internal signing service.
  • Controls bypassed. They triggered transfers outside the normal withdrawal process, skipping the withdrawal service, risk checks and approvals.
  • Keys protected, transactions not. MPC protected the private keys; it did not stop a trusted service from signing the wrong transaction.

The drain

  • First transfers: XRP began moving at 20:05 UTC. Within about 20 minutes, hot wallets on ten chains were empty.
  • Largest sweep: One Ethereum transaction at 20:24 moved $2.53 billion — 187,464 ETH and 63 tokens, including 1.6B USDT.
  • Last transfer: The final unauthorized transfer, 2,961.9 ETH, settled at 23:56.
  • Why pausing wasn’t enough: The attackers had used EIP-7702 to delegate an Ethereum hot wallet to their own contract. Pausing withdrawals did not revoke that permission, so funds kept leaving until Binance cleared the delegation.

Aftermath

  • Out of reach: About $3.49 billion was out of reach. Bitcoin, ETH, XRP and SOL have no issuer that can freeze them.
  • Following the money: The attackers swapped roughly $760 million of stablecoins into ETH and TRX before the freezes landed, at a 12–20% loss, then split the Bitcoin across 19 addresses and moved Tron funds through bridges.
  • Attribution: Elliptic and Chainalysis linked the laundering to the September Bitget theft; the FBI attributed the campaign to North Korea.
  • What was frozen: Issuers froze $2.85 billion, 94% of it stablecoins. Tether froze $1.57B and Circle $823M; Ripple, Paxos, First Digital and World Liberty Financial also froze their tokens.
  • Chain-level response: Arbitrum and Optimism governance moved stolen assets out of reach.
  • Recovery: Binance refilled the wallets, absorbed the loss, and reopened withdrawals on 1 November at 06:39 UTC. Users lost nothing.

Lessons for everyone

  • 🧰 Don’t install untrusted third-party tools. A package that looks normal can replace a trusted binary with a same-name copy.
  • 🎭 Treat job interviews as an attack surface. Fake recruiters, fake assignments and fake repos are standard social engineering now.
  • 🧠 Train people, not just systems. The first click was human; the controls around it were not enough. Make it easy to question unexpected requests and report them without blame.
  • 🔑 Scope credentials tightly. A deployment token on a laptop turned a workstation breach into a path to production.
  • 🛑 Pausing is not revoking. On-chain permissions live outside your infrastructure. Design the emergency path before you need it.

Humans can still be the weakest link. Alongside technical weaknesses, attackers exploit human trust. Companies should invest in understanding how social engineering is evolving—not just patching code. Verify recruiters through independent channels, run unfamiliar assignments in isolated environments, and give staff realistic, recurring training backed by controls that limit the damage when someone is deceived.

Sources: Binance initial findings, 3 Nov 2026; Elliptic; FBI; Chainalysis.