Operational security threats and controls

The largest digital-asset losses of recent years did not come from broken cryptography. They came from people being deceived, insiders being bribed, software being tampered with and signers approving what they could not read. Knowing how these attacks work, with real cases and figures, is the starting point for designing controls that hold up.

The threat landscape

Case study: Bybit, February 2025

On 21 February 2025 the Dubai-based exchange Bybit lost about 401,000 ETH, worth roughly USD 1.5 billion at the time. It is the largest crypto theft on record. On 26 February 2025 the FBI attributed it to North Korean actors it tracks as "TraderTraitor," also widely known as the Lazarus Group.

The target was Bybit's Ethereum cold wallet, a Safe multisig that needed several signers. The attack unfolded in steps:

1. Weeks earlier, the attackers compromised the computer of a developer at Safe{Wallet}, the company behind the wallet's web interface.
2. With that access, they replaced a JavaScript file of the Safe{Wallet} web application, hosted in Amazon Web Services, with malicious code. The code was written to activate only for Bybit's wallet, so other users saw nothing unusual.
3. When Bybit's signers started what looked like a routine transfer from cold to warm storage, the interface showed them the expected transfer. The data actually sent for signing was different: a delegatecall that swapped the wallet's contract logic for code controlled by the attackers.
4. The signers approved it on their hardware wallets, which showed raw data instead of a readable description. This is blind signing.
5. With control of the wallet, the attackers drained it. The malicious code was removed from the website minutes later.

Investigators found no compromise of Bybit's own infrastructure, and the Safe smart contracts themselves were not flawed. The multisig worked exactly as designed: enough signers approved. The weak point was the path between what the signers intended and what they signed.

Two earlier cases followed the same pattern. In July 2024 the Indian exchange WazirX lost about USD 230 million from a multisig wallet; according to WazirX, the transaction its signers approved did not match what their custody provider's interface displayed. In October 2024 the lending protocol Radiant Capital lost about USD 50 million after malware on developers' devices made their hardware wallets sign malicious transactions while the interface showed legitimate ones.

Lessons for any custodian:

• Verify transactions on an independent channel, not on the same interface that prepared them.
• Block high-risk transaction types (delegatecall, owner changes, contract upgrades) by default in the policy engine.
• Treat every software provider in the signing path, including wallet front ends, as a critical supplier.
• Refuse to sign anything the device cannot decode.

Phishing and social engineering

Well-funded attackers rarely attack keys head-on. They target the people around them. Common lures include fake job offers with "technical tests" that install malware, messages from people posing as former contractors, fake meeting links, and deepfake voice or video calls. The Radiant attack reportedly began with a malicious file sent by someone posing as a former contractor.

Controls:

• Phishing-resistant multi-factor authentication (FIDO2 hardware security keys) for every privileged system
• Dedicated signing devices, never used for email, browsing or messaging
• Out-of-band verification of any unusual instruction through a known contact
• Training based on real attack patterns, not generic awareness slides

Insider risk

Insiders can do damage without ever touching a key. In May 2025 Coinbase disclosed that criminals had bribed overseas customer-support contractors to copy customer data, including names, addresses, government ID images and account balances, for less than 1% of monthly transacting users. The attackers then demanded USD 20 million, which Coinbase refused to pay. The stolen data was used to impersonate Coinbase and trick customers into sending funds. Coinbase estimated the cost of remediation and reimbursements at USD 180–400 million.

Controls: least privilege (support staff see only what each case needs), monitoring of bulk data access, background checks, rotation of sensitive roles, quick removal of access, and quorums so that no single insider can move funds.

Supply chain attacks

Custody platforms depend on libraries, browser extensions, cloud services and vendors. In December 2023 an attacker phished a former Ledger employee and published a malicious version of the Ledger Connect Kit, a JavaScript library used by many decentralized applications. For a few hours it displayed a fake prompt that drained wallets, and about USD 600,000 was stolen. In September 2025 an npm maintainer was phished, and widely used packages such as chalk and debug were briefly replaced with versions that swapped crypto addresses in web pages.

Controls: pin dependency versions and review updates, keep a software bill of materials (SBOM), use reproducible builds and subresource integrity for web code, run security assessments of vendors in the signing path, and keep a verification step that does not depend on the same software.

Address poisoning

In an address poisoning attack, the attacker generates an address whose first and last characters match one the victim uses, then sends a zero-value or tiny transaction so the look-alike appears in the victim's history. Victims who copy an address from their history send funds to the attacker. In May 2024 one holder lost 1,155 WBTC, about USD 68 million, this way. Most of it was returned after negotiation, which is rare.

Controls: allowlists, never copying addresses from transaction history, checking the full address and not just its ends, and wallet software that flags look-alikes.

SIM swaps and account takeover

In a SIM swap, the attacker persuades or bribes a mobile carrier to move the victim's number to a new SIM card, then receives the victim's SMS codes and password resets. In January 2024 the US SEC's account on X was taken over after a SIM swap, and a false post announcing approval of spot bitcoin ETFs moved the market. Controls: no SMS-based authentication for anything important, carrier port-out locks, and phone numbers used for recovery that are kept private.

Physical coercion

Wrench attacks, where criminals use threats or violence to force someone to transfer assets, have become more frequent. In January 2025 David Balland, a co-founder of the hardware-wallet maker Ledger, was kidnapped in France and later freed by police. Further kidnappings and attempts targeting crypto entrepreneurs and their families were reported in France during 2025.

Controls: design so that no single person can move material funds, time locks that make quick transfers impossible, duress procedures, discretion about who holds keys and how much, and personal security for exposed executives.

A control framework

NIST Cybersecurity Framework 2.0 (2024) organizes controls into six functions: govern, identify, protect, detect, respond and recover. For custody, they translate as follows:

Exam focus: the Bybit case (21 February 2025, about 401,000 ETH or USD 1.5 billion, compromised Safe{Wallet} developer machine, malicious interface code, blind signing of a delegatecall, FBI attribution to North Korea's TraderTraitor) and its lessons; WazirX and Radiant as similar interface attacks; phishing-resistant MFA and dedicated signing devices; insider risk and the May 2025 Coinbase bribery case; supply chain attacks (Ledger Connect Kit, npm 2025); how address poisoning and SIM swaps work; controls against physical coercion; organizing controls with the NIST CSF 2.0 functions.