The Binance Hack, Separated Into Facts, Assessments and Guesses

DAte

Category

Blockchain

Reading Time

7 Min

The Binance Hack, Separated Into Facts, Assessments and Guesses

Binance says no private key leaked. Eleven of its hot wallet keys still signed transfers to the attacker. Both statements can be true, and that is where the analysis starts.

Below is the October 29 incident in three layers: what is confirmed, what analysts assess, and what is still a guess.


The numbers


Binance puts the impact at $6.34 billion. TRM Labs counts $6.36 billion, Chainalysis $6.4 billion, Elliptic $6.43 billion. The spread reflects pricing at different moments.

Funds left hot wallets on Bitcoin, Ethereum, Tron, Solana, the XRP Ledger, Arbitrum, Avalanche, Optimism, Base and Polygon. Binance says it replenished 98% of affected assets by October 31 from its own funds. Withdrawals began returning on November 1.

About $2.8 billion is frozen. Elliptic estimated $3.24 billion still sat in theft linked addresses on November 3.


The entry path, per Binance


Binance published its account on November 4. Nobody has independently verified the early stages.

June. A DevOps engineer on the wallet team installed a tampered Homebrew installer while setting up an AI coding assistant. Binance caught it, replaced the device, reset credentials. The attackers kept one thing: knowledge of who ran the wallets.

September. A fake recruiter approached an engineer. Two interview rounds ended with a test assignment run on a work computer. It replaced a legitimate developer tool with a malicious one. Endpoint protection missed it.

About nine days before the theft. From that machine, through an internal AI tool, the attackers reached two more wallet staff computers. One held a saved access token allowing new software to run on Binance systems.

October 29, after 17:00 UTC. A script used that access to call the internal wallet signing service and approve transfers outside the normal withdrawal process.

How the AI tool was abused is undisclosed. So are the "other breaches" Binance mentioned on October 30. The pattern fits what we described in Beyond Human Hackers: The Rise of AI Driven Attacks in Crypto.


The execution


TRM's onchain timeline, UTC:

  • 20:05 to 20:09. Five XRP Ledger wallets send 233 million XRP, each left at minimum balance.

  • 20:20. A Bitcoin hot wallet sends 18,270 BTC, split into 1,000 and 100 BTC chunks in the same block. A Tron wallet sends 612.7 million USDT.

  • About 20:23. The main Ethereum hot wallet is swept: 167,464 ETH plus 1.6 billion USDT.

  • Until 23:56. Repeat sweeps catch newly arriving deposits.

It was rehearsed. The operator wallet was funded through Tornado Cash 36 hours earlier. Code was tested on two wallets, dry run on a third, then deployed.


Why Ethereum fell differently


The Ethereum wallet was not drained by a normal transfer. The attacker used EIP 7702, shipped in the Pectra upgrade in May 2025. One signed authorization lets a wallet delegate execution to a smart contract. The delegation persists until the wallet signs a new one.

Authorizations are bound to the wallet's nonce, which advances with every withdrawal. The attacker submitted 50 authorizations for 50 consecutive nonces. One matched. The wallet now answered to a sweeper contract.

This is why the bleeding lasted four hours. Binance shut down withdrawal systems, but the delegation lived onchain. The wallet used multiparty computation, so no single key could sign a quick revocation. Engineers had to improvise a procedure. Background on delegated wallet logic: Account Abstraction: Making Crypto Usable for Normal Humans.


Containment and laundering


Tether froze the main attacker wallet 37 minutes after the Ethereum sweep. Circle froze 823 million USDC. Ripple, World Liberty Financial and the Arbitrum Security Council froze smaller amounts.

The attacker anticipated this. About 325 million USDT was swapped into ETH through intent based contracts at an effective price more than 10% worse than market. Paying that to exit a freezable asset is rational.

Chainalysis reports highly automated movement since. On November 9, Elliptic described a new route: single use memecoins on NEAR. Buyer wallets pay into a liquidity pool, separate seller wallets withdraw. No direct transfer links them, so address blocklists miss the payout. Volume through this route is small so far, about $430,000.


What is certain


  • Roughly $6.3 to 6.4 billion left Binance hot wallets within about four hours.

  • The transfers carried valid signatures from Binance's own keys.

  • The Ethereum drain used EIP 7702 delegation.

  • About $2.8 billion is frozen.

  • Binance has covered the losses.

Attribution to TraderTraitor, the North Korean unit behind the 2025 Bybit theft, is a high confidence assessment from TRM and Chainalysis. The FBI named the same group. Chainalysis's strongest evidence is stolen funds commingling in real time with proceeds from an earlier Bitget theft. North Korea denied involvement on November 3.


Two hypotheses


These go beyond the evidence.

1. The failure was authorization, not custody. Binance says no key leaked. Blockaid says eleven keys signed. The consistent reading, also raised by Check Point, is that the attacker controlled what the signer was asked to sign. If so, the MPC setup did its job. It protected keys. It had no way of knowing whether a request came from the real withdrawal pipeline. Our guess: withdrawal policy was enforced upstream of the signer, and a direct call to the signer bypassed it.

2. The 50 authorizations point to a narrow window. Banking 50 Ethereum signatures in advance only makes sense if the attacker could not count on signing at execution time. That suggests access was time limited or request gated. If correct, monitoring the volume and shape of signing requests, not only where funds go, could have flagged the staging before the theft. TRM raises this possibility without committing to it.


The human failure


Strip away EIP 7702 and MPC, and the chain starts with one decision. An engineer, interviewing for another job, ran a stranger's test assignment on his work computer. That machine sat one hop from wallet staff whose access moved billions.

A serious lapse. The wallet team was targeted in June, so the threat was known. Job hunting from a corporate laptop inside the custody perimeter breaks the most basic rule of this work: nothing untrusted runs on a device that can reach the people who sign.

The organization shares blame. A laptop with that reach should not execute outside code freely, and a token able to deploy software should not sit saved on a workstation. For teams running hot wallets, two questions follow. Does your signer enforce withdrawal policy itself, or trust whatever calls it? And can one employee's bad afternoon still reach it? More in How Blockchain Systems Fail in the Real World (and How to Prevent It).

Conclusion

The wallet team was targeted in June, so the threat was known. Job hunting from a corporate laptop inside the custody perimeter breaks the most basic rule of this work: nothing untrusted runs on a device that can reach the people who sign.

Photo of article author
Dariusz Wróbel

CEO

CEO with 15+ years in software engineering, bridging deep technical expertise with real-world business execution.

Share post

Related News

Related News

Diagram showing The Binance Hack, Separated Into Facts, Assessments and Guesses

A layered analysis of the $6.3B Binance hack: the confirmed attack chain, the EIP 7702 drain, laundering routes, and two hypotheses about what failed.

Diagram showing The Binance Hack, Separated Into Facts, Assessments and Guesses

A layered analysis of the $6.3B Binance hack: the confirmed attack chain, the EIP 7702 drain, laundering routes, and two hypotheses about what failed.

Diagram showing Vibe Coding Smart Contracts: Where AI Helps and Where It Fails

AI can speed up smart contract development, but vibe-coded contracts fail in predictable ways. Where generated code breaks and where AI genuinely helps.

Diagram showing Vibe Coding Smart Contracts: Where AI Helps and Where It Fails

AI can speed up smart contract development, but vibe-coded contracts fail in predictable ways. Where generated code breaks and where AI genuinely helps.

Diagram showing Designing Orokai: The Non-Custodial Tradeoffs Nobody Talks About

How Base58 designed Orokai's non-custodial app: signing flows, token approval defaults, seed phrase onboarding, and where security and UX directly conflict.

Diagram showing Designing Orokai: The Non-Custodial Tradeoffs Nobody Talks About

How Base58 designed Orokai's non-custodial app: signing flows, token approval defaults, seed phrase onboarding, and where security and UX directly conflict.

Diagram showing Security Token vs Utility Token: Which Do You Need?

How the SEC's 2026 taxonomy and MiCA actually classify tokens, why the security vs utility binary is obsolete, and how to design a token that holds up.

Diagram showing Security Token vs Utility Token: Which Do You Need?

How the SEC's 2026 taxonomy and MiCA actually classify tokens, why the security vs utility binary is obsolete, and how to design a token that holds up.

>

>

The Binance Hack, Separated Into Facts, Assessments and Guesses