On October 29, attackers used a Binance service that signs cryptocurrency transactions to make unauthorized transfers. Binance says approximately $6.34 billion in digital assets were affected. The FBI later attributed the theft to North Korea. [1][2]
By the time Binance suspended withdrawals, the attackers had already established permissions that could operate outside its normal withdrawal system. Shutting that system down did not cancel them. [1]
How the attackers got there
Binance's November 3 report traces the attack back months before the transfers. In June, an engineer downloaded tampered developer software while trying to install an AI coding tool. Binance says it detected the problem and replaced the device, but the attackers learned which employees worked on its wallet systems. [1]
In September, another engineer was approached through a fake job interview and asked to run a coding assignment. The assignment contained malicious software. By October 20, the attackers had used that compromised computer and an internal AI tool to reach two other engineers' systems. They found a credential that allowed software deployments and, on October 29, used stolen access to run malicious code that reached Binance's wallet-signing service. [1]
The attackers moved from employee computers into internal deployment access and then into a service that signs wallet transactions. Binance has not released the internal records needed to independently reconstruct every step, including precisely how the AI tool was used. Its report does not establish that the AI model independently decided to move money or that prompt injection was involved.
Why stopping withdrawals wasn't enough
A cryptocurrency transaction can carry a valid digital signature even when the company responsible for the wallet never approved the transfer. The network can verify that signature. It cannot check the company's internal approval process.
Binance says the attackers used its signing service and an Ethereum feature called EIP-7702 to give a program they controlled permission to act for one of its wallets. EIP-7702 is an intended feature that allows this kind of delegation. It was the unauthorized use of Binance's signing authority, not proof that Ethereum's signature system was broken. [1][3]
Those permissions were recorded on Ethereum. They remained active after Binance suspended normal withdrawals and shut down affected services. The exchange says its last unauthorized transfer finalized at 23:56 UTC on October 29. It cleared the malicious delegation at 00:10 UTC on October 30. [1]
A separate published analysis of Ethereum transactions identifies the delegation and its removal. It could not establish who requested the signatures inside Binance or whether the wallet's underlying private keys were stolen. [4]
There is a familiar version of this problem in ordinary operations. An internal tool can grant access to another system. Disabling the original tool might not remove the access it already issued. That permission has to be identified and revoked where it is still valid.
What this means for automated work
The same question comes up when software can move money, change account permissions, send information or act on behalf of a team. Having the credentials to complete an action does not establish that the action was approved.
That's something I pay attention to as I build operational systems. I want to know who requested the action, whether it was within their authority, what actually changed and whether the permission can be revoked. Those checks need to work when something goes wrong, not just when the workflow behaves as expected.
In an October 27 blog post, NIST researchers proposed design tenets for AI-agent identity. Those tenets call for clear agent identities, limited delegated permissions and access that can be revoked. That guidance predates the Binance attack and is not an assessment of its cause. [5]
What is still unresolved
As of its November 3 report, Binance said approximately $2.85 billion in affected assets had been frozen or intercepted. It reported that 98% of affected token assets had been replenished as of October 31, and withdrawals were reopening in phases. Frozen assets and replenished holdings are not the same thing as funds independently verified as recovered. Binance's investigation and outside security review were still ongoing. [1][6]
We can follow the reported attack without treating an early investigation as a complete explanation of how every internal control failed.
A system can accept a transaction with a valid signature without knowing whether the organization intended to authorize it.
Sources
- Binance, November 3, 2026. 29 October 2026 Binance Security Incident: Initial Findings and Response Update. Preliminary company investigation.
- FBI, October 31, 2026. North Korea Responsible for $6 Billion+ Binance Attack. Official law enforcement attribution.
- Ethereum, EIP-7702: Set Code for EOAs. Protocol specification.
- GCSA, October 30, 2026. Independent EIP-7702 Transaction Link Analysis. On-chain analysis only, not Binance internal forensics.
- NIST, October 27, 2026. Design Tenets for Agent Identity Management Systems. Proposed guidance.
- Binance, November 3, 2026. Recovery Update #6: Withdrawals Reopen for More Assets. Service restoration status.