StatewardStateward
← Toutes les vulnérabilités
CRITICALWeb3exploited in the wildcurated

WEB3-DRIFT-2026

Web3 · Solana · Drift Protocol

Résumé

On 1 April 2026, attackers drained about $285 million from Drift, the largest perpetual-futures exchange on Solana, in roughly twelve minutes, more than half of everything the protocol held. There was no smart-contract bug. The contracts worked exactly as written. The attackers, later tied to North Korea's Lazarus operations, spent months socially engineering the people with privileged access, then abused a Solana feature called durable nonces to get the protocol's multisig signers to approve, in advance and without realising it, the transactions that handed over control. Once they had admin rights they whitelisted a worthless token as collateral, deposited a pile of it, and borrowed out the real money. It is the case that proves the modern crypto heist is an operations and social-engineering problem, not a Solidity one.

How it happened

From the autumn of 2025, the attackers ran a long con. Posing as a quantitative trading firm courting the project, they built trust with Drift contributors and worked their way into cloud infrastructure and, critically, into the trust of the people who controlled the protocol's privileged multisig. The technical key was Solana's durable nonces, a legitimate feature that lets a transaction be signed now and submitted validly much later, instead of expiring within seconds like a normal Solana transaction. Using that, the attackers got the Security Council multisig members to sign transactions that looked routine but were actually pre-authorising an administrative takeover, to be executed at the attackers' chosen moment. A migration with no timelock then stripped the remaining safeguards instantly. With admin control, they whitelisted a worthless token of their own as acceptable collateral, deposited a large amount of it, and withdrew roughly $285 million in real assets, USDC, SOL, and ether, bridging it to Ethereum within hours.

Who was behind it

Blockchain-intelligence firms attributed the theft to North Korea with medium-to-high confidence, and Drift's own post-mortem tied it to the actors behind the October 2024 Radiant Capital hack, a Lazarus-adjacent cluster tracked as UNC4736 (also called AppleJeus or Citrine Sleet). It is the same national operation, and often the same playbook of patient social engineering, behind Bybit, the Ronin bridge, and, going back further, Sony Pictures and WannaCry. The trend it confirms is stark: by 2026 the most expensive crypto attacks are no longer clever contract exploits but social engineering against the humans and processes that hold privileged access.

The aftermath

The funds were bridged out and converted quickly, and no recovery was confirmed; most of the $285 million was gone. Drift's total value locked fell from around $550 million to under $300 million almost immediately, making it the largest DeFi theft of 2026 at the time and the second largest in Solana's history after the $326 million Wormhole bridge hack. The deeper damage was to a comfortable assumption: that a multisig with several independent human signers is a strong control. Here it was the control that got captured, because the signers were manipulated into approving the takeover themselves.

Why Drift still matters

Drift shows that privileged access is the prize, and people and process are the weak points. The contracts were fine; the governance and signing process around them was not. The defences are operational, not cryptographic: treat every signer's workstation and the whole signing workflow as a high-value target; never blind-sign, and have signers independently decode and simulate exactly what each transaction does before approving; be deeply suspicious of any mechanism, like durable nonces, that lets a signature be executed later than expected, and constrain or monitor it; require timelocks on administrative and migration actions so a takeover cannot complete instantly and can be caught and cancelled; and train contributors against the long, patient social-engineering campaigns Lazarus specialises in. The lesson rhymes with Bybit: the keys were never the weak link, the humans approving with them were.

Comment le corriger

  • Treat the signing and governance infrastructure as fully compromised: revoke and rotate all privileged keys and rebuild the multisig with signers verified on clean systems.
  • Trace and publicly flag the stolen funds immediately, and coordinate with exchanges and bridges to freeze what is still reachable.
  • Remove any standing administrative path that lacks a timelock, so no single approval can hand over control instantly.

Comment l’éviter

  • Put timelocks on all administrative, migration, and collateral-whitelisting actions, so a malicious change is delayed long enough to be detected and cancelled.
  • Mandate clear-signing: signers must independently decode and simulate each transaction on isolated devices, and never blind-sign a pre-built payload.
  • Constrain and monitor durable nonces and any feature that lets a signed transaction execute later than expected.
  • Harden signer endpoints and contributor accounts against long-running social engineering, and verify out-of-band any party requesting privileged interaction.
  • Require multiple independent reviews of decoded calldata before any privileged transaction is approved.

Références

Vulnérabilités liées

Tout Web3 →