Skip to content
Crypto.world

The Crypto Industry

The Keys Were Safe: Four Crypto Thefts at the Signing Step

Bybit, KelpDAO, Bitget and Drift were robbed without a stolen private key or a smart-contract bug. Their own records show where each theft happened instead, in what signers and verifiers were shown…

The Keys Were Safe: Four Crypto Thefts at the Signing Step

Four Thefts and the Step Each Attacker Corrupted

Between February 2025 and September 2026, attackers took about $1.5 billion from Bybit, 116,500 rsETH from the KelpDAO bridge, about $388 million from Bitget and about $295 million from Drift, by the victims' or investigators' own counts. In none of the four did those accounts find a stolen private key or a bug in a smart contract. Each attacker corrupted a step before the key was used: what signers saw or were asked to approve, what a verifier read from a node, or what a signing service was told to send.

The keys signed what they were given, so custody's weak point has moved from key storage to the authorization pipeline around the key. So "cold storage", "multisig" and "audited contracts" no longer tell a customer how safe a custodian is: Bybit's loss came out of a multisig cold wallet, and Drift's contracts held. The four records point to narrower controls instead: checking what is being signed independently of the screen that proposes it, delaying privileged actions long enough for someone to notice, and requiring more than one independent verifier.

Incident Loss, as the record gives it What the attacker corrupted What held Record
Bybit, 21 Feb 2025 About $1.5 billion (FBI, 26 Feb 2025) JavaScript served to the signers by Safe{Wallet}, disguising the transaction Signers' keys; Safe smart contracts FBI; Safe; Sygnia
KelpDAO rsETH bridge, 18 Apr 2026 116,500 rsETH, about $292 million at the time (LayerZero Labs, May 2026) Two internal RPC nodes read by the only required verifier Verifier signing keys; on-chain checks LayerZero Labs; Kelp
Bitget, 24 Sep 2026 About $388 million, final figure (Bitget, October 2026) Wallet back-end, reached with credentials taken through a third-party security product Private keys; cold wallets Bitget; Mandiant; SlowMist
Drift, 1 Apr 2026 $295 million in user funds (Drift, 5 May 2026) Two of five Security Council approvals, pre-signed with durable nonces Programs and contracts; no evidence of seed-phrase compromise Drift; on-chain record

Bybit: Three Signers, One Disguised Transaction

On 21 February 2025, Bybit moved funds from one of its Ethereum multisig cold wallets to a warm wallet, a routine transfer approved through Safe{Wallet}'s web interface. The transaction the signers approved at 14:13 UTC was not the one on their screens. Bybit's co-founder and chief executive, Ben Zhou, described what happened on X about an hour and a half later:

It appears that this specific transaction was musked, all the signers saw the musked UI which showed the correct address and the URL was from @safe . However the signing message was to change the smart contract logic of our ETH cold wallet. — Ben Zhou, Co-founder and CEO, Bybit

Sygnia, one of two firms whose preliminary reports Zhou published on 26 February, found the cause in the browser caches of all three signers' machines. JavaScript files served from app.safe.global through an Amazon S3 bucket had been modified on 19 February with code that activated only for two contract addresses, one of them Bybit's, and rewrote the transaction. The signed payload delegated the wallet's logic to a pre-deployed malicious contract whose sweep functions moved over 400,000 ETH without further approvals. About two minutes after the transaction, the injected code was removed.

On 28 February, the Safe Ecosystem Foundation published the Safe{Wallet} team's finding that the attack "was achieved through a compromised Safe{Wallet} developer machine resulting in the proposal of a disguised malicious transaction." External researchers, it said, found no vulnerabilities in the Safe smart contracts or in the source code of the front end and services. Excerpts from Mandiant's preliminary report, posted by Safe{Wallet} and summarized by Sygnia on 16 March 2025, date the first compromise to 4 February. A developer's macOS workstation ran a Docker project likely delivered through social engineering.

The attackers then worked inside Safe{Wallet}'s AWS account. Their attempt to register their own MFA device failed, so they used the developer's active session token instead. The FBI's public service announcement of 26 February 2025 (I-022625-PSA) said North Korea was responsible for the theft of approximately $1.5 billion, activity the bureau calls "TraderTraitor." Bybit's own count, published on 3 March 2025, was $1.46 billion, mostly 401,347 ETH.

KelpDAO: One Verifier, Two Poisoned Nodes

The KelpDAO rsETH bridge, built on LayerZero's cross-chain messaging protocol, lost 116,500 rsETH on 18 April 2026, about $292 million at the time of the attack by LayerZero Labs' count. In this design, a contract on the destination chain releases tokens once the verifiers the application requires, called decentralized verifier networks (DVNs), attest that a matching message was sent on the source chain. Each application chooses its own DVNs. Kelp's statement the next day said a forged packet was accepted, releasing the rsETH "from the Ethereum bridge lockbox to an attacker-controlled address."

LayerZero Labs' incident report, published in May 2026, traces the intrusion to 6 March. A developer on the team that ran LayerZero's RPC nodes was socially engineered into cloning a malicious GitHub repository, which installed two backdoors, FLATROOF and ROOFDECK, on a Mac. Single sign-on and mandatory MFA were enforced, and the endpoint detection software did not flag the malware. From 30 March, the attacker used the developer's harvested session keys to enter LayerZero's cloud.

Between 16 and 18 April, the attacker patched the running node software of two internal RPC clusters in memory. The patched nodes gave LayerZero's monitoring tools correct answers and gave the DVN forged ones. At 16:30 UTC on 18 April, a denial-of-service attack on external RPC providers forced the DVN's failover logic onto the two poisoned nodes, and at 17:35 UTC the rsETH was unlocked. The report's layer-by-layer review finds the DVN's private keys uncompromised and the on-chain quorum check intact:

The signatures the contract verified were produced by the legitimately-held keys, signing the hash they were given. — LayerZero Labs, KelpDAO Incident Report

One compromised verifier was enough because of the channel's configuration. The report says "a previous 2-of-2 configuration had been modified by the application owner to a 1-of-1 configuration which used only the LayerZero Labs DVN," so no second verifier could disagree and reject the forged packet. Kelp's statement of 19 April placed the compromise "outside of Kelp's operational domain," in infrastructure "operated by a third-party provider." The two accounts assign the failure to different parties; both describe a forged message that passed verification.

CrowdStrike and Mandiant attribute the attack to North Korea with high confidence and, with medium confidence, to UNC4899, also known as TraderTraitor, which the report names as the group behind the Safe{Wallet} attack at Bybit. LayerZero says its DVN now refuses to sign for any channel where it is the only required verifier and reads security data through a quorum of independent RPC sources. LayerZero Labs has also raised the protocol's defaults: it completed the update of Endpoint v2's default DVN configurations to a minimum 3-of-3 in July 2026 and changed Endpoint v1's default message library on 4 August, and on 7 August it said: "All applications using default message libraries on active pathways are now secured by a minimum 3-of-3 configuration." LayerZero said applications with pinned libraries would not be affected.

Bitget: Forged Withdrawal Commands Below the Risk Checks

At 18:31 UTC on 24 September 2026, unauthorized transfers began leaving part of Bitget's hot and warm wallet infrastructure. At 19:05, its reconciliation system detected the discrepancy and its risk controls automatically blocked withdrawal requests across the platform. Containment began at 19:40, and at 21:44 the wallet withdrawal services, signing services included, were shut down. Bitget's notice that evening estimated the loss at approximately $351.6 million; by October its incident page gave a final verified figure of approximately $388 million, from 12 hot or warm wallet addresses on 11 blockchains.

In Bitget's account, a zero-day vulnerability in a third-party security product gave the attackers access to critical internal management systems and high-privilege credentials. The forged identity information appeared valid, so "the systems treated the access as legitimate internal activity." The attackers reached wallet back-end servers and injected forged withdrawal commands, which were processed as normal withdrawals, bypassing risk-control verification before withdrawal records were generated. They then deleted traces of the commands. Bitget's summary: "Private keys were not compromised, and cold wallets were unaffected."

The investigators' reports, released on 30 September, add detail. Mandiant's status report of 28 September describes a web shell on one of two third-party security appliances and lateral movement to the production wallet job server. SlowMist traces malicious activity in the logs back to the end of August and describes a recovered withdrawal tool "tailored to the wallet system’s withdrawal logic" that forged risk-control parameters. Its on-chain count puts the transfers between 18:31 and 21:23 UTC, about 2 hours and 52 minutes.

The platform-wide block on withdrawal requests at 19:05 did not end the theft: transfers continued for more than two hours, consistent with commands that entered below the layer that handles requests. At 20:40, because private-key compromise had not yet been ruled out, Bitget began moving funds to cold wallets; the investigation later ruled it out. Withdrawals fully resumed on 2 October 2026.

Bitget's written record names no attacker. Its chief executive, Gracy Chen, said in a live broadcast on X hours after the theft that some IP addresses matched the VPN choices of a North Korean group; she also said the attacker's identity was not confirmed. On 28 September The Block reported that she still suspected the same group but would not name it before the incident report. Bitget Academy warns that "Public speculation regarding attacker attribution should only be treated as confirmed if supported by verified investigative findings."

Drift: Approvals Signed Days Before They Were Used

Drift, a perpetual-futures exchange on Solana (since renamed Velocity DEX), lost control of its protocol on 1 April 2026. Its first technical account is the preliminary thread it posted on X on 2 April. There, Drift said the attack "was not the result of a bug in Drift’s programs or smart contracts" and that there was "no evidence of compromised seed phrases." It described "unauthorized or misrepresented transaction approvals obtained prior to execution, likely facilitated through durable nonce mechanisms and sophisticated social engineering."

A normal Solana transaction is rejected once its recent blockhash is more than 150 slots old. A durable nonce removes that expiry, which suits multisigs whose approvals arrive at different times, and lets a signature given today be used days later. Per Drift, four durable nonce accounts were created on 23 March, two associated with Security Council members and two with the attacker. After a planned council change on 27 March, another followed on 30 March, and the attacker again had "effective access to 2/5 signers in the updated multisig."

The on-chain record matches Drift's sequence. A test withdrawal from the insurance fund executed at 16:04:16 UTC on 1 April. At 16:05:18 and 16:05:19 UTC, four slots apart, two transactions that open with the nonce-advance instruction marking a durable-nonce transaction created, approved and executed a change of the protocol's admin. With admin control, Drift says, the attacker introduced a malicious asset and removed all pre-set withdrawal limits.

On 5 April Drift followed with an "Incident Background Update", still described as preliminary, that links the intrusion to a roughly six-month approach by people presenting as a quantitative trading firm, whom contributors met in person at industry conferences. Among what it believes may have been three attack vectors, it names a code repository one contributor may have been compromised after cloning and a TestFlight application a second contributor was induced to download. With medium-high confidence, it assessed the operation as the work of UNC4736, the North Korean group Mandiant had linked to the October 2024 Radiant Capital hack.

Drift's first estimate was approximately $280 million withdrawn; its recovery plan of 5 May 2026 put the loss at $295 million in user funds. That plan calls the durable-nonce attack surface "central to how the April 1 exploit was executed" and says it has been removed. Drift's 3 June update says Mandiant "conclusively attributed" the attack to UNC6862, which Drift describes as a North Korean threat group. The account here rests on Drift's own statements and the chain's record, not on a published investigator's report.

What to Score Instead of "Cold Storage" and "Multisig"

The Safe and Drift smart contracts, a bridge's on-chain checks and an exchange's private keys all held, yet attackers took control of a multisig cold wallet and of a two-of-five council's administrative powers. In each case the attack arrived as an approval, an attestation or a command that the system accepted as genuine. A due-diligence review that stops at where keys are stored grades the part of the system these attackers left alone. Four questions cover the part they used.

Is the transaction checked outside the screen that proposes it? Bybit's signers saw the correct address on a page served by Safe{Wallet}'s own, compromised infrastructure. Drift's framework for its relaunch requires dedicated signing devices, "with transaction content independently verified outside the primary signing interface before any signature is executed." In February 2025, Safe committed to lead an industry-wide initiative to "increase verifiability of transactions." A custodian should be able to name the tool that decodes each payload and who runs it.

Do privileged actions wait? Drift's admin change went through in two transactions one second apart; Bitget's transfers ran for close to three hours, much of it after detection. Drift's relaunch framework sets timelocks "on all critical administrative actions, paired with real-time alerts," and disables durable nonces for all signers. A delay helps only if someone is watching it, so the score is the length of the window and who receives the alert.

How many independent parties must agree? A second required DVN, LayerZero's report says, "would have required the attacker to coordinate simultaneous attacks on multiple independent DVN signers, off-chain infrastructure or RPC dependencies." Independence covers data as well as keys: a verifier that reads one set of nodes is one point of failure, however many keys it holds. The protocol still enforces whatever an application configures, including a 1-of-1 stack with a different DVN, so the verifier set is something to inspect rather than assume.

Who can reach the signing path, and for how long? The Safe{Wallet} and LayerZero intrusions both ran on a developer's live session despite MFA; Bitget's ran on credentials that appeared valid. Drift's 5 April update names a cloned code repository and a TestFlight app as possible ways in, both offered to its contributors by people they had known, and met in person, for months. It asked other teams to "audit who has access to what, and treat every device that touches your multisig as a potential target."

LayerZero says privileged access now requires just-in-time elevation with short-lived credentials, multi-person review for IAM changes, and device and session validation on every administrative request. Bitget says multiple approvals are now required for critical operations, and that it is tightening controls on third-party products.

Parts of the record are still open. Bitget has said a formal security report with fuller technical detail will follow, and Mandiant's 28 September status report described its investigation as ongoing. LayerZero said on 15 October 2026 that the admin roles of its Solana Endpoint and ULN302 library are moving to OneSig, its one-click signing tool, with the migration due to be completed by mid-November.

Sources

  1. FBI IC3, PSA I-022625-PSA: North Korea Responsible for $1.5 Billion Bybit Hack (26 February 2025) ic3.gov
  2. Safe Ecosystem Foundation, Statement by the Safe Ecosystem Foundation (28 February 2025) safefoundation.org
  3. Bybit, Bybit Security Incident: Timeline of Events and FAQs (3 March 2025) bybit.com
  4. Ben Zhou, Bybit co-founder and CEO, on X (21 February 2025) x.com
  5. Ben Zhou on X: Bybit Hack Forensics Report, Sygnia and Verichains preliminary reports (26 February 2025) x.com
  6. Sygnia, Bybit: What We Know So Far (16 March 2025) sygnia.co
  7. LayerZero Labs, KelpDAO Incident Report (May 2026) layerzero.network
  8. LayerZero, LayerZero Labs KelpDAO Incident Report, blog (20 May 2026) layerzero.network
  9. Kelp DAO, Incident Response (19 April 2026) blogs.kerneldao.com
  10. Bitget, Security Notice: Bitget exchange hot wallets incident (24 September 2026) bitget.com
  11. Bitget, Security Incident: Official Progress Update (incident page) bitget.com
  12. Bitget Academy, Bitget Hack Explained: Timeline, Impact and Security Response bitget.com
  13. Bitget, Update on SlowMist and Mandiant Report (30 September 2026) bitget.com
  14. Mandiant, Bitget Incident Response Status Report (28 September 2026) img.bgstatic.com
  15. SlowMist, Investigation Progress Report: Bitget (September 2026) raw.githubusercontent.com
  16. Cointelegraph, Bitget Suspects North Korea Behind $352M Hack (25 September 2026) cointelegraph.com
  17. Decrypt, Bitget Hack Losses Climb to $387M: Here’s What Happened, and Why North Korea Is a Suspect (25 September 2026) decrypt.co
  18. The Block, Bitget attacker tested risk controls with small transfers before $388 million theft, CEO says (28 September 2026) theblock.co
  19. Drift Protocol on X, incident thread, part 1 (2 April 2026; account since renamed Velocity) x.com
  20. Drift Protocol on X, incident thread: findings and first estimate (2 April 2026) x.com
  21. Drift Protocol on X, incident thread: what the attacker was able to do (2 April 2026) x.com
  22. Drift Protocol on X, incident thread: durable nonce timeline (2 April 2026) x.com
  23. Drift Protocol on X, incident thread: April 1 execution phase (2 April 2026) x.com
  24. Drift Protocol on X, incident thread: postmortem to follow (2 April 2026) x.com
  25. Drift Protocol on X, Incident Background Update (5 April 2026) x.com
  26. Drift, Incident Recovery Update (16 April 2026) drift.trade
  27. Drift, Recovery Plan for Affected Users (5 May 2026) drift.trade
  28. Drift, Drift Recovery Update (3 June 2026) drift.trade
  29. Solana on-chain record: Drift insurance-fund test withdrawal (1 April 2026, 16:04:16 UTC) solscan.io
  30. Solana on-chain record: durable-nonce admin-transfer proposal (1 April 2026, 16:05:18 UTC) solscan.io
  31. Solana on-chain record: durable-nonce admin-transfer execution (1 April 2026, 16:05:19 UTC) solscan.io
  32. Solana documentation, Transactions (blockhash expiry) solana.com
  33. Solana cookbook, Durable and Offline Transaction Signing using Nonces solana.com
  34. LayerZero, Ecosystem Update (15 October 2026) layerzero.network
  35. LayerZero Labs, Ongoing Security Updates (4 June 2026) layerzero.network
  36. LayerZero Labs, Ongoing Security Updates 7/10 (10 July 2026) layerzero.network
  37. LayerZero Labs, Ongoing Security Updates 8/7 (7 August 2026) layerzero.network

Companies in this story

Related stories

Explore

More articles