← Breach analysis by the Dual Vulnerabilities framework

Radiant Capital $50M multisig takeover (Citrine Sleet, 2024)

2024 · Hong Kong · K - Financial and insurance activities · Theft via compromise of multisig signers' devices

Analysed 2026-11-06 by Hermes

Open this case in the tool →

What happened

Theft via compromise of multisig signers' devices

Dual vulnerability type

Strategy to Execution

Evidence coverage: 4/6 stages, weight 23/60

Decisive stage: response (clearest evidence) · Alternative: Expertise to Execution

Why this classification

(1) Radiant Capital is a cross-chain DeFi lending protocol whose entire treasury sat behind an Admin multisig, and its security doctrine was formally real: industry-standard signing SOPs (Tenderly simulation, payload verification, hardware wallets) and a contracted, pre-configured Automated Incident Response programme with Hypernative alerting. (2) In practice, on 11 September 2024 a Telegram ZIP from an actor impersonating a former contractor - its domain spoofing the contractor's site - installed the INLETDRIFT backdoor on Radiant developer Macs; five weeks later the malware intercepted the multisig signing flow so signers approved benign-looking transactions while a transferOwnership was signed behind the scenes, the attackers seized the LendingPool contracts and drained about $50M on Arbitrum and BSC; the automated response fired and auto-paused the pools, but pausing does not stop an ownership takeover - the BSC markets were drained 80 seconds after the pause and compromised owners were removed only 4h27m later. (3) The classification rests on the response stage: the strategic expectation existed as a decided and resourced programme (doctrinal origin), yet the operational playbooks and signing architecture did not realise it for the access-control scenario (execution break) - formal artefacts unhinged from operational reality, the Strategy to Execution diagonal. This is a documented gap, not generic weak security: the controls existed and demonstrably did not translate. The entry vector (impersonation malware) is load-bearing for the theft, but the entry and containment gaps - unverified inbound files and signing on ordinary developer laptops - are readings the victim's own disclosure contradicts in part (it maintains the SOPs were followed and the compromise was undetectable at the signing stage), so they point to Expertise to Execution only by inference and never win the decisive-stage rule; a secondary Expertise to Execution pattern remains visible in the distribution. (4) RM36 (Strategy to Execution lens): the incident-response plan existed and fired, but it assigned no executable response to an access-control compromise - only a pause action that the attacker's upgrade path ignored, with owner removal taking 4h27m (causal role: Constraining). RM04 (Strategy to Execution lens): architecture concentrated control of the whole cross-chain treasury in one signing layer whose members used ordinary developer laptops, so one class of endpoint compromise captured the 3-of-11 threshold and the blast radius was every market (causal role: Amplifying). RM24 (Strategy to Execution lens): verification requirements for inbound documents from external contributors were not translated into working procedure - a ZIP whose domain spoofed the contractor's site circulated unchecked among several developers (causal role: Enabling). RM34 (Strategy to Execution lens): monitoring requirements were implemented on-chain but not on the attack surface that mattered - weeks of signing-failure anomalies reported by the signers themselves were never handled as compromise indicators and the 35-day implant went undetected (causal role: Obscuring).

Evidence by stage

Risk markers

Sources

Notes

Link check: medium.com blocks automated access (HTTP 403); the page was opened and read during research. Link check: blockonomi.com was unreachable at check time (timeout); the page was opened and read during research. Description mismatch (minor): the EuRepoC text says the attackers 'gained access to three private keys' by 'exploiting smart contract vulnerabilities'; per the victim's own post-mortem the private keys stayed on Ledger/Trezor hardware wallets - the attackers compromised the signers' computers and intercepted the signing flow so valid signatures were produced for a malicious payload (a blind-signing / device-compromise pattern, confirmed by the Mandiant-led update). Amounts vary by source: ~$50M (Radiant post-mortem), ~$51.5M (EuRepoC title), $48M + $5-6M via infinite user approvals (Hacken). Attribution: Mandiant via Radiant's 6 December 2024 incident update, high confidence, DPRK-nexus UNC4736/Citrine Sleet (also tracked as AppleJeus); BleepingComputer confirms. FBI involvement reported by Blockonomi (18 Oct 2024), matching the case file's legal_response fields. Disclosure: first public warning came from De.Fi on X (case file source_disclosure: 'Incident disclosed by IT-security company'); Radiant published its post-mortem 17 Oct and the Mandiant summary 6 Dec. Ancilia, a third-party security firm, inadvertently posted a wallet-drainer link while assisting victims (assodigitale) - response confusion at a third party, not a victim-side gap, so not coded. Related row 2728 (case file opened): a separate earlier incident, the 2 January 2024 flash-loan exploit ($4.5M via a rounding error); the weakness is not comparable to the multisig signer compromise, so RM21 is not coded - noted only as a prior incident at the same victim within 2024. Country taken from eurepoc_extra.receiver_country (Hong Kong); Radiant operates as a DAO and the jurisdiction was not independently verified - not guessed beyond the dataset field. Losses were never meaningfully recovered.