The Binance Exploit: Why Signing Authority Is the New Security Perimeter
The largest digital asset theft on record did not need a stolen key. It needed a signing service that would approve anything, and that is the lesson for every regulated institution entering digital assets.
In this article

On 29 October 2026, attackers moved about $6.34 billion in digital assets out of Binance hot wallets across ten blockchains. Most of it left in roughly twenty minutes. The FBI has attributed the attack to TraderTraitor, a North Korean state-sponsored group, making it the largest theft of its kind on record.
The headline number is not the most useful part of the story. Binance’s findings point not to a stolen private key but to unauthorised access to an internal wallet signing service, which the attacker used to approve transactions that never passed through the normal withdrawal process. The keys stayed protected. The authority to use them did not.
That distinction matters to every regulated financial institution and fintech that holds, moves or settles digital assets, including stablecoins and payment tokens. Most security budgets are built around protecting keys. This incident shows the perimeter has moved to the systems that decide what those keys sign.
Twenty minutes, ten chains
According to Binance, the first unauthorised transfers began at 20:05 UTC, when five XRP hot wallets sent 233.46 million XRP to an attacker-controlled address. Solana, Optimism and Bitcoin wallets followed. On Ethereum, the attacker used the signing service to produce 50 signing transactions that set up malicious EIP-7702 delegations, giving an attacker-controlled smart contract execution rights over a Binance hot wallet.
Once those delegations were on-chain, suspending withdrawals no longer stopped the outflow. The delegations worked independently of Binance’s withdrawal infrastructure, so the attacker could keep sweeping assets, including new deposits that the exchange’s own consolidation system routed into the compromised wallet. Independent research from Four Pillars shows that a single sweep moved more than 60 types of assets worth about $2.53 billion, and that the delegation code was visible on-chain for 3 minutes and 24 seconds before that sweep ran. The last unauthorised transfer finalised at 23:56 UTC; Binance cleared the delegation at 00:10 UTC on 30 October.
One detail deserves attention from anyone running multi-party computation (MPC). Because the hot wallet used MPC, there was no single private key that could immediately sign a revocation. Binance had to build and run a special procedure, which added time. MPC did its job of keeping the key whole and safe. It also made the emergency response slower.
The months before the twenty minutes
The theft took minutes. The preparation took months. Binance’s timeline describes a campaign that combined social engineering, supply-chain malware and internal tooling:
- June 2026. A wallet DevOps engineer searched for a popular AI coding tool and installed it through a tampered installer. The device was caught and replaced and credentials were rotated, but the engineer’s role and team had been exposed.
- September 2026. A software engineer was approached on a legitimate job platform for a fake role. After two convincing interview rounds, he was asked to download an “assignment” to a work computer. It carried previously unseen malware that endpoint tools did not detect.
- 20 October 2026. From that computer, the attacker used an internal AI tool to take control of two wallet engineers’ machines and found a token with container deployment permissions.
- 29 October 2026. The attacker deployed a container with the stolen credentials and ran the script that drove the signing service.
Binance says it detected and stopped 57 suspected North Korea-linked attempts during 2026 before this one succeeded. That is the uncomfortable arithmetic of a persistent, well-funded adversary: defenders have to win every time.
An AI tool that can run commands on an engineer’s machine and reach other systems is, in effect, another insider with broad privileges.
AI appears twice on the attack path: once as bait, once as the bridge between machines. As regulated firms roll out coding assistants and internal agents, the permissions those tools hold belong on the same access reviews as any privileged employee.
Keys were safe. Signing was not.
The pattern is not new. Four Pillars traces the same shape through earlier exchange losses attributed to North Korea-linked groups: in DMM Bitcoin (2024), WazirX (2024), Bybit (2025) and Bitget (2026), the attackers manipulated what was signed, or the screen that showed signers what they were approving, rather than stealing a key outright.
MPC and hardware security modules are designed to stop keys from leaking. They do not decide whether a request is a legitimate withdrawal, a delegation authorisation or something else entirely. If the signing layer does not read and check every request against policy, a compromised internal service can ask the key to sign anything. An exchange’s withdrawal operations never need an EIP-7702 delegation; a policy engine that understood that would have refused all 50.
Containment came from across the ecosystem
Much of the damage was contained by parties outside the exchange. Tether and Circle froze about $2.4 billion in stablecoins within roughly an hour, and chains and token projects used administrative powers to freeze or recover more. By 3 November, Binance put the total frozen or intercepted at about $2.85 billion, and said 98% of affected assets had been replenished from its own funds by 31 October.
Customer confidence largely held. Reuters reported net withdrawals of $4.72 billion in the days after the exchange reopened, about 3% of assets on the platform, with $2.7 billion of inflows over the same period. Few firms could absorb a loss of this size from their own resources. For smaller platforms, the lesson is that the capacity to make customers whole is part of the security model, not separate from it.
There is also a harder lesson in the freezes. Attackers now know exactly which assets and routes can be frozen. Blockchain analytics firm Elliptic reported that the hackers still controlled about $3.28 billion, mostly in bitcoin and ether, which cannot be frozen by an issuer. The next attacker will convert freezable assets faster.
What the UAE’s rulebook already asks for
The UAE anticipated this risk. Dubai’s Virtual Assets Regulatory Authority (VARA) requires licensed virtual asset service providers to cover not only key generation and storage but “the signing and approval of transactions” in their technology governance and risk framework. The same rulebook requires no single point of failure in access to virtual assets, strict access controls with audit logs, and procedures to revoke a signatory’s access immediately.
Those requirements map directly onto the failure points in this incident: a signing service that approved requests outside policy, privileged tokens that sat on an engineer’s machine, and authorisations that took hours to revoke. For institutions building digital asset, stablecoin or payment token capabilities in the UAE, the framework is already pointing at the right place. The work is in implementing it to the standard this attack sets.
A playbook for anyone who signs
- Put a policy engine in front of every key. The signer should decode each request, classify it (withdrawal, delegation, contract call, approval) and refuse anything outside an allow-list, however it arrived.
- Ban transaction types you never use. If operations never need delegations, token approvals or arbitrary contract calls from a hot wallet, block them at the signing layer.
- Rehearse revocation before you need it. Pre-sign or pre-approve emergency revocation transactions, and test how long MPC or multi-signature revocation actually takes.
- Watch your own wallets on-chain. Alert in seconds on code attached to a hot wallet, new approvals or unusual counterparties. Every on-chain change is public the moment it lands.
- Treat AI tools as privileged identities. Give internal assistants and agents least-privilege access, separate them from signing and deployment credentials, and log what they do.
- Harden hiring and developer channels. Brief engineers on fake recruitment, keep personal assignments off work devices, and verify developer tools against signed sources.
- Plan the balance sheet response. Know in advance how losses would be covered, how reserves would be proven, and how withdrawals would reopen in phases.
The bottom line
The largest digital asset theft on record did not require a stolen key. It required months of patient access-building and one signing service that would sign whatever it was asked. For regulated financial institutions and fintechs entering digital assets, protecting keys is the minimum. The control that matters now is the one that decides what those keys are allowed to sign, and how fast that permission can be taken back.
- Reflections on the Binance Exploit — Four Pillars Research, November 2026
- 29 October 2026 Binance Security Incident: Initial Findings and Response Update — Binance, 3 November 2026
- North Korea Responsible for $6 Billion+ Binance Attack — FBI, 31 October 2026
- Binance clients pull $4.7 billion after hack as crypto world studies fallout — Reuters, 2 November 2026
- Binance seeks to shore up confidence after record $6.3 billion heist — Reuters, 30 October 2026
- FBI says North Korea behind $6 billion Binance crypto theft — DW, 31 October 2026
- Technology and Information Rulebook, D. Cryptographic Keys and VA Wallets Management — Virtual Assets Regulatory Authority (VARA), Dubai
Ripae helps regulated financial institutions and fintechs design, build and modernise payment systems, from real-time rails and wallets to cross-border flows.


