Commentary · Cyber incident response · Fraud & asset recovery
Beyond private keys: lessons from the US$6.34bn Binance security incident
- On 29 October 2026, Binance reported about US$6.34 billion of impacted assets from its hot wallets, with about US$2.85 billion frozen or intercepted by 3 November. Both figures are provisional.
- No private key was stolen, and MPC didn't fail. Attackers worked through staff computers, a deployment credential and an internal AI tool until Binance's own signing system moved the funds.
- Protecting keys is only part of the job. Platforms must control who and what can instruct the signing system, by any route.
- Boards should be able to prove, with tests and timings, that unauthorised transfers fail and that incidents can be contained on-chain.
On 5 November 2026, the Blockchain Association Singapore's Web3 Cybersecurity Sub-Committee, which I co-chair, published a 20-page preliminary industry advisory on the Binance incident. Its key contributors were Marc Krisjanous, Michael Lew, Fadli and Thomas Wan. It draws on Binance's own initial findings and other public sources. It is an independent industry assessment, not a regulatory notice, a finding that Binance's controls were deficient, or a forensic report. Here are the lessons I think matter most, and why they matter legally.
What happened, according to Binance's preliminary findings
- June 2026. A wallet-team engineer looking for an AI coding tool through a search engine downloaded a tampered copy of Homebrew, a common developer installer. Binance replaced the computer and changed credentials, but the engineer's role and colleagues had already been exposed.
- September. A second engineer was approached through a legitimate job platform by a fake "project director", interviewed by a fake HR representative and CEO, and told to download a coding "assignment" onto a work computer. It installed previously unknown malware.
- 20 October. Using that computer and an internal AI tool, the attackers took control of two wallet engineers' computers, one of which held a deployment credential.
- 29 October. The attackers used stolen credentials to deploy a malicious program that reached the wallet-signing system. Funds left hot wallets across the XRP Ledger, Solana, Optimism, Bitcoin and Ethereum.
- The standing instruction. The attackers also used the signing system to authorise an EIP-7702 delegation on an Ethereum hot wallet, effectively a standing instruction recorded on-chain. Suspending withdrawals didn't cancel it. Revoking it needed MPC signers, a valid on-chain transaction and a specially developed procedure.
The FBI attributed the theft to North Korean "TraderTraitor" activity on 31 October. How the internal AI tool was misused has not been publicly disclosed.
The five lessons
1. Developer endpoints and deployment pipelines need key-level protection
The route to the signer ran through ordinary staff computers, a fake recruitment approach and a deployment credential. Allow only approved software on work devices, keep job-interview code off them, keep deployment credentials off staff computers, make them short-lived, and require a second person to approve changes to live systems.
2. The signing system must check every request, whatever the route
MPC protects key shares. It doesn't decide what should be signed. The signer acted on instructions from outside the normal withdrawal process. Every signing request should be checked against independently protected rules bound to the exact transaction, unsupported transaction types and delegations should be refused by default, and aggregate outflow limits should apply across wallets and chains.
3. Incident response has to reach the blockchain
Halting withdrawals internally didn't stop the on-chain delegation. Platforms need chain-specific runbooks to revoke delegations and approvals, stop new signing, and move residual assets to safe wallets, and should rehearse them, including with a degraded MPC quorum.
4. Treat AI tools as privileged users
An AI tool that can read code, run commands or reach other systems acts with whatever access it is given. Give each tool narrow, read-only-by-default access, enforce limits outside the model rather than in prompts, keep it away from signers, log what it does, and plan how to switch it off. This fits squarely with the new MAS AI risk management guidelines.
5. Speed of reporting drives recovery
By 3 November, about US$2.85 billion had been frozen or intercepted with law enforcement, analytics firms and industry partners. Stablecoin issuers can only freeze tokens at addresses they know about, so how quickly a platform reports and shares receiving addresses affects how much is frozen. Note too that Binance replenishing about 98% of affected assets from its own resources is a different thing from stolen assets being recovered or frozen. The two figures shouldn't be combined.
Why this matters legally
- Evidence of reasonable care. After a loss, users, counterparties, insurers and regulators will ask whether the platform's controls were reasonable. The advisory's "evidence of effectiveness" approach (negative tests, measured detection and halt times, documented access reviews) is exactly the record a platform will need. See who carries the loss when platforms are breached.
- Board accountability. The advisory expects management to show boards that prohibited transfers and deployments fail in controlled tests, how much is exposed in hot wallets, and timed evidence of detection and containment.
- Recovery and insurance. Fast notification to issuers, exchanges and law enforcement, and preserving claims against vendors and intermediaries, determine what can be recovered. See protect, report, recover.
- Regulatory mapping. The advisory notes that controls should be mapped, entity by entity, to the MAS Technology Risk Management Guidelines and the Cyber Security Agency's TR 153:2026.
Frequently asked questions
Was a private key stolen?
Why didn't suspending withdrawals stop the losses?
Want your incident response plan tested against these scenarios? Get in touch, or read the first 72 hours.
Credits. This piece is based on Binance Security Incident: Lessons for Digital Asset Platforms and Custodians, a preliminary independent industry cybersecurity advisory by the Blockchain Association Singapore Web3 Cybersecurity Sub-Committee (Version 1.0, 5 November 2026). Its key contributors were Marc Krisjanous (BAS Fellow, Zanarc), Michael Lew (Co-Chair, BAS Web3 Cybersecurity Sub-Committee), Fadli (BAS Fellow) and Thomas Wan (BAS Secretariat). The analysis, diagrams and recommendations are theirs. The "Why this matters legally" section is my own.
Figures and findings are preliminary, as at 5 November 2026, and attributed to Binance or the sources cited in the advisory. This is general information, not legal advice, and not a finding about any party's compliance.
