Back to all articles

Binance $6.3B hack: valid signatures, EIP-7702 and failed controls

Share

Updated 4 November 2026: Binance's 3 November report adds the internal intrusion chain and forensic findings. Original publication date: 2 November 2026; URL unchanged.

On 29 October 2026, Binance hot wallets sent approximately $6.3 billion in cryptoassets to attacker-controlled addresses, using cryptographically valid signatures. This is what makes the incident so instructive: a valid signature is not evidence that the business decision requesting that signature was legitimate.

On 31 October, the FBI attributed the theft to North Korea's TraderTraitor group. Independent investigations by TRM Labs and Elliptic describe a multi-chain operation, including a particularly revealing abuse of Ethereum's EIP-7702 delegation mechanism.

Binance says it has not found evidence so far that its wallet private keys were leaked. That statement is not definitive proof that all possible key exposure has been ruled out. But the observable transfers are consistent with an attacker obtaining signatures through compromised trusted infrastructure, without breaking the underlying cryptography.

Confirmed timeline as of 2 November 2026

Official disclosures and independent on-chain findings provide the following chronology:

  • 29 October, around 20:05 UTC: unauthorized withdrawals begin across multiple Binance hot wallets. Binance estimates approximately $6.3 billion lost; TRM estimates $6.36 billion and Elliptic $6.43 billion, reflecting different scopes and valuations.
  • Around 20:20 UTC: an Ethereum transaction contains 50 EIP-7702 authorizations signed by a Binance hot wallet, each pointing to an attacker-controlled contract.
  • Around 20:24 UTC: after an authorization takes effect, the contract triggers a massive asset transfer. The attacker repeatedly sweeps newly deposited funds for hours.
  • 30 October: Binance confirms the incident, temporarily suspends withdrawals and revokes affected authorizations. Its investigation initially mentions a compromised DevOps laptop, but subsequently reveals additional points of compromise.
  • 31 October: the FBI formally attributes the attack to TraderTraitor, publishing addresses linked to illicit movement of assets.
  • 1 November: Binance begins a phased withdrawal resumption.
  • 3 November: the official report identifies $6.34 billion in affected assets and $2.85 billion frozen or intercepted. Binance says 98% of the impacted token assets had been replenished as of 31 October, which is not the stolen-funds recovery rate.

Replenishing exchange reserves and freezing stolen tokens are different processes. The first covers a financial shortfall; the second restricts the movement of particular assets. Neither, on its own, means the exchange has definitively recovered all stolen funds.

Updated 4 November: Binance discloses the intrusion path

Binance's 3 November initial report describes a months-long operation. These are Binance's preliminary findings, pending independent expert review.

June: a fake Homebrew installer exposed the wallet team

A Wallet DevOps engineer (A) searched for an AI coding tool and installed a tampered Homebrew package. Binance detected the anomaly, replaced the computer and rotated related credentials. But the attacker had already learned the engineer's role and the identities of wallet-team colleagues. Rebuilding a device cannot erase intelligence about personnel already acquired by an adversary. Binance does not claim credentials revoked in June were reused in the final attack.

September: fake job interviews and malicious code assignment

Another engineer (B) was approached via a genuine recruitment platform. Following two fake interviews, including a video meeting with a supposed HR representative, the employee received a coding assignment. Malware in the repository ran during setup.

Binance describes replacing a legitimate pyenv component at /opt/homebrew/Cellar/pyenv/2.8.0/libexec/pyenv, modifying symlinks and executing zsh. The substituted component could run commands and contact an attacker-controlled server. According to Binance, endpoint tools failed to detect it at the time. These details are case-specific macOS forensic leads, not universal detection rules.

20 October: an internal AI tool used for lateral movement

Using B's compromised workstation, the attacker exploited access through an internal AI tool to reach environments belonging to A and a third wallet engineer (C). A container deployment permission token was discovered on A's machine.

No prompt injection, inherent language-model flaw or autonomous AI-led attack is established by the report. It documents an internal tool used during lateral movement, rather than an AI independently planning an intrusion.

29 October: container deployment and access to the signer

Between 17:15 and 19:56 UTC, the intruder accessed C's machine, used credentials stolen from A, deployed a container and ran malicious code. Binance says this enabled unauthorized access to an internal wallet signing service and operations outside the normal withdrawal process. The supported sequence is compromised endpoints → lateral movement → deployment permission → malicious script → valid signatures.

Why pausing withdrawals did not revoke EIP-7702

EIP-7702 delegation exists on-chain, even when a service's usual withdrawal workflow is disabled. Stopping customer withdrawals alone therefore could not revoke the attacker's delegation.

The affected hot wallet used multi-party computation (MPC), meaning no single ready-to-use private key could produce an immediate revocation. Binance says a special procedure was required. The last unauthorized transfer it identified (2,961.9 ETH) finalized at 23:56 UTC on 29 October, after having remained pending. At 00:10 UTC on 30 October, Binance sent an authorization from Binance14 to the zero address to clear the delegation. This does not prove a cryptographic weakness in MPC; it demonstrates why emergency on-chain revocation procedures must be prepared and tested.

Forensic evidence and defensive controls

In an authorized investigation, preserve browser and download history, Homebrew installation records, pyenv and symlink changes, shell artifacts, fake recruitment communications, internal AI tool access and invocation logs, deployment token histories, container image digests, and wallet signer audit records.

For custody environments, separate request, approval, deployment and signing, limit internal AI-tool permissions and rehearse EIP-7702 revocation. The SANS analysis of fake Homebrew installers is a comparable but separate sample, not an attribution to Binance's malware.

How can a wallet sign a malicious transaction?

A cryptographic signature proves that the correct key authorized a transaction under the chain's rules. It does not prove which human or application requested it, what the human saw, or whether the workflow sending the signing request was legitimate.

Consider a simplified production withdrawal flow:

  1. An application decides that a withdrawal should be made.
  2. An independent policy layer checks identity, permissions, destination and limits.
  3. A signing service signs the transaction.
  4. The blockchain verifies the signature and executes the transaction.

If an attacker compromises the authorization layer or obtains a route around expected approvals, the cryptographic verification at step four may work flawlessly even though the operation is fraudulent.

In its 30 October incident update, Binance said a compromised DevOps engineer's laptop was an early investigative lead, but subsequent work uncovered more intrusion points. A comprehensive root-cause report is not yet public. The blockchain alone cannot reveal the internal path by which the signing requests were approved.

EIP-7702: legitimate account functionality, malicious authorization

EIP-7702, “Set Code for EOAs”, introduced with Ethereum's 2025 Pectra upgrade, lets a regular externally owned account (EOA) delegate its account code to a smart contract. The account's key signs the authorization; the blockchain cannot infer why the authorization was produced.

This creates legitimate programmable wallet capabilities. EIP-7702 is not itself a software vulnerability. The risk is that an account may sign an authorization pointing to hostile code or to a delegate with insufficient guardrails.

According to TRM's transaction reconstruction, the attacker submitted 50 previously signed authorizations associated with the hot wallet known as Binance 14. All targeted an attacker-controlled sweeper contract, but each contained a different consecutive nonce.

The nonce is the account's transaction counter. Because this wallet was processing many withdrawals, its count was moving. A batch of 50 consecutive nonces ensured that one authorization could match the counter at the moment the transaction landed.

One valid authorization was enough. The wallet began executing the attacker's delegated code. About three and a half minutes later, one transaction moved 167,464 ETH and dozens of tokens, including about 1.6 billion USDT. The delegation persisted, so newly incoming funds could be swept again without obtaining a new signature for every subsequent transfer.

The crucial lesson is that a single signed authorization can change what an account will do in future transactions. This was not just a malicious one-off transfer.

What the other chains reveal

TRM and Elliptic also identified illicit transfers on Bitcoin, Tron, the XRP Ledger, Solana and other networks. Most of these were ordinary validly signed transfers, not EIP-7702 operations.

That pattern supports the inference that the attacker reached transaction request and signing infrastructure across chains, rather than exploiting a cryptographic flaw in Ethereum. It remains an analyst assessment, not a final root cause published by Binance.

Some investigations differ on the number of chains they count because of variations in scope, network classification and transaction attribution. The important observation is that multiple chain-specific signing systems produced valid transfers to the attacker.

Digital forensics: what can the evidence actually show?

A defensible investigation combines on-chain artifacts with off-chain operational evidence. Each answers different questions.

On-chain evidence

Investigators can preserve and examine:

  • 0x04 set-code transactions, EIP-7702 authorization lists and associated signatures;
  • delegated contract addresses, deployed bytecode, transaction and call traces;
  • the evolution of account nonces and delegations, later sweeps and subsequent deposits;
  • transfers between chains, DeFi swaps, issuer freezes and blocklisting actions;
  • block times, event ordering and execution traces, without confusing observable sequence with proof of causality.

The EIP specification defines SET_CODE_TX_TYPE = 0x04 and the signed authorization tuple. Researchers can therefore independently check the delegation mechanism using raw transaction evidence rather than relying solely on press statements.

Off-chain evidence

Investigators also need API gateway logs, signer and HSM audit records, container and workstation artifacts, cloud identity events, orchestrator changes, deployment records, approval workflows and historical policy configuration.

A valid signature shows which key was used, not who was authorized to instruct the signer. Off-chain evidence is necessary to establish how a request reached the signing boundary, whether policy checks were bypassed and whether any secrets were exposed.

Preserve source provenance, accurate timestamps and appropriate cryptographic hashes during authorized investigations. The FBI's published address list provides leads for tracing and screening; it should not be treated as a complete or permanently fixed list of every related address.

Controls that can contain a signing-infrastructure compromise

A custody architecture must secure the decision to sign, not only the key:

  1. Separate request, approval and signing. A compromised DevOps identity, service or container must not be enough to authorize high-value withdrawals.
  2. Enforce independent transaction policy. Amount and velocity limits, destination allowlists, supported chain IDs and asset-specific controls should be enforced outside the requesting application.
  3. Review the real effect of an authorization. An EIP-7702 authorization changes future account behavior. Verify delegated destinations, contract code and effective permissions, ideally with independent transaction simulation.
  4. Monitor unexpected account delegation. Detect changes in EIP-7702 delegate code, unknown implementation addresses and repetitive sweeps of new deposits.
  5. Reduce correlated multi-chain exposure. One control-plane identity should not be able to trigger unrestricted transactions across all custody systems.
  6. Rehearse emergency response. Provide ways to halt signing and withdrawals while retaining tamper-resistant evidence, and test these controls in advance.

No architecture can guarantee zero risk, but these measures make it harder for one compromised identity or application workflow to generate valid, high-impact signatures at scale.

Frozen funds are not the same as replenished assets

By 30 October, Binance reported that around $3 billion was frozen on-chain. Independent analyses from TRM, Bitquery and Elliptic provide evolving figures around $2.7–$2.85 billion, depending on which assets are counted, prices and measurement time.

A token freeze means specific assets cannot currently be moved because an issuer or an emergency mechanism can restrict them. It does not necessarily mean ownership or custody has returned to Binance.

Binance's 3 November report says the exchange has set aside more than 100% of the funds necessary to cover losses and replenished 98% of impacted token assets as of 31 October. More illiquid tokens were still being acquired. The 98% figure is not the share of stolen funds recovered, nor necessarily an equal percentage of total US-dollar value. Binance now reports $2.85 billion frozen or intercepted out of $6.34 billion in affected assets. Freezing, replenishment and actual recovery are different processes.

Lessons for GLOV Secure and GLOV Consulting

For GLOV Secure, the practical lesson is that self-custody, institutional storage and multisig security are not simply matters of selecting keys or hardware devices. Approval workflows, trusted transaction displays, delegated authority, independent policy enforcement and monitoring are all security boundaries.

Our earlier analysis of valid signatures and ownership evidence in the Qubic incident explores a different question. Here, the signature may be perfectly valid while the business instruction was malicious. Our piece on AI agents, cryptographic signing and the limits of human authorization addresses a similar separation between requesting an operation and approving it.

For GLOV AI, still an initiative in preparation, this case illustrates the need to audit internal AI tool permissions and logs without assuming unproven autonomous attacker capabilities.

For GLOV Consulting, the key issues are privileged access, service-to-service trust, transaction approval audits and multi-chain resilience. This complements our analysis of verification failures in cross-chain bridges, while recognizing that the Binance case is not a bridge vulnerability.

Binance customers should not enter wallet seed phrases on unsolicited “incident verification” sites or pay fake recovery services. Withdrawal availability and account-related instructions must be verified through the exchange's official channels.

Evidence gaps and uncertainties

Binance's 3 November report now documents a preliminary intrusion chain and unauthorized access to its signing service. It remains a first-party initial account, with the announced independent review still unpublished. Precise internal AI tool capabilities, complete malware functionality and the signer policy checks bypassed remain partly unknown. Its statement that it has found no evidence of a private-key leak is not definitive proof that every possible leak has been excluded.

EIP-7702 authorizations and on-chain transfers are observable, but do not by themselves identify the specific employee, internal process or system fooled by the attacker. Financial totals vary as prices, freezes, transfers and reporting scopes change.

The TraderTraitor attribution comes from the FBI's 31 October statement and threat intelligence; it does not identify every individual operator.

GLOV Solutions editorial analysis — originally published 2 November 2026, updated 4 November 2026. No compromise of GLOV Solutions or Snow-Fall is alleged.

Official and independent sources

Related articles

Back to all articles