Radiant Capital Multisig/Device Compromise (Oct 16, 2024)

Date: 2024-10-16
Loss: ~$50M
Category: DeFi exploit (signer-device compromise / blind-signing)
Attack vector: Malware on multiple signers' devices that displayed benign transactions in the front end while malicious transactions were signed in the background
Attribution: UNC4736 / Citrine Sleet (DPRK-aligned, assessed sub-cluster of Lazarus) per Mandiant

Overview

On October 16, 2024, Radiant Capital, a cross-chain lending protocol operating on Arbitrum and BNB Chain, was drained of approximately $50 million in a sophisticated attack that, like the later Bybit cold-wallet compromise, exploited the gap between what signers saw in their front end and what their devices actually signed. The attackers, attributed by Radiant's incident-response partner Mandiant to UNC4736 (also tracked as Citrine Sleet), a DPRK-aligned threat actor assessed as a sub-cluster of the Lazarus apparatus, had compromised the devices of multiple Radiant core contributors via malware delivered, per Radiant's post-mortem, through a Telegram message from someone impersonating a trusted former contractor and bearing a malicious file disguised as a project document. When the Radiant team conducted a routine multisig operation, the malware presented legitimate-looking transaction data in the Safe (Gnosis Safe) front end while the hardware devices were caused to sign malicious transactions in the background, allowing the attackers to harvest enough signatures to seize control of contracts and drain funds. Radiant suspended its lending markets the same day. Despite an 18-month recovery effort, the protocol was unable to recover a meaningful portion of the stolen funds or raise sufficient fresh capital, and in June 2026 Radiant Capital announced it would wind down. The incident is a pivotal precursor to Bybit in the canon of blind-signing failures: it demonstrated, at the protocol level, that audited Safe multisigs can be defeated when the signers' own machines are compromised and the displayed transaction diverges from the signed transaction.

Timeline of events

Radiant Capital's post-mortem dates the intrusion's origin to September 11, 2024, when a Radiant developer received a Telegram message from an actor impersonating a trusted former contractor, asking for feedback on a supposed new project and attaching a ZIP file. The file, shared among several developers for review, delivered malware that compromised multiple core contributors' devices. With that foothold established over the following weeks, the attackers waited for an opportunity to subvert a legitimate multisig operation. On October 16, 2024, during a routine transaction (a scheduled adjustment to the protocol's parameters that required multisig approval), the malware on the compromised devices manipulated what was signed: the Safe front end displayed the expected, benign transaction to each signer, while the underlying hardware wallets were induced to sign malicious transactions that transferred control. By harvesting the required signatures, the attackers obtained the authority to take over Radiant's lending-pool contracts and drained approximately $50 million across Arbitrum and BNB Chain; reporting noted that the attackers moved roughly $52 million as they consolidated and bridged the proceeds in the days that followed. Radiant suspended its lending markets on October 16 to halt further damage. Mandiant was engaged for incident response and subsequently attributed the attack to UNC4736 / Citrine Sleet. Over the following 18 months Radiant attempted to recover funds and raise capital; on June 1, 2026, it announced it would wind down, unable to recover meaningfully or secure a viable runway.

Attack mechanism: blind-signing at the protocol level

The technical heart of the Radiant attack is the same class of failure that would later define the Bybit cold-wallet compromise: a divergence between the transaction a human signer believes they are approving and the bytes the signing device actually signs. Radiant used a Safe (Gnosis Safe) multisig to govern its protocol, requiring multiple signers to approve administrative transactions, a standard, audited, widely-trusted setup. But the security of a multisig assumes that each signer is reviewing and approving the genuine transaction. The attackers defeated that assumption by compromising the signers' devices with malware: when the team initiated a routine multisig operation, the front-end interface displayed the legitimate, benign transaction the signers expected, while the malware caused the hardware wallets to receive and sign different, malicious transactions in the background. The signers, seeing the expected data on their screens and following their normal procedure, approved what they believed were routine changes; in reality they were authorizing the attacker's takeover. Because enough signers were compromised to meet the multisig threshold, the attackers accumulated the signatures needed to control the protocol's contracts and drain the pools. This is blind-signing, the human verifies the front end rather than the actual payload on a trusted device, applied to a DeFi protocol's governance multisig, and it is precisely the trust-boundary gap (between rendered UI and signed bytes) that the Bybit post-mortem would later make the central lesson of 2025.

Root cause analysis

Three compounding root causes produced the loss. First, the initial compromise came from the Lazarus-family fake-contractor/recruiter social-engineering vector: a malicious file delivered via a trusted-seeming Telegram contact, run by developers, compromising their machines, the same playbook seen in the DMM Bitcoin (vendor employee) and broader DPRK campaigns. Engineering teams handling privileged keys are high-value targets, and unsolicited files, even from seemingly known contacts, are a live malware vector. Second, and most fundamental, the signing process relied on the signers visually verifying the transaction in the Safe front end rather than independently verifying, on a trusted out-of-band device, the actual transaction payload (the calldata and its decoded effects) before approving. Because the front end could be manipulated by device malware, visual verification of the UI was not a reliable control. Hardware wallets that clearly decode and display the true transaction effects, and an independent verification step comparing the computed transaction hash on a separate clean machine, would have exposed the divergence. Third, multiple signers were compromised, defeating the multisig's core assumption that signers are independent and that compromising the threshold is hard; if signers share devices, software, or a common malware-delivery vector, the multisig's independence is illusory. The unifying lesson, identical to Bybit's, is that an audited multisig is only as secure as the integrity of each signer's device and the fidelity of what that device displays.

Response, wind-down and the Bybit parallel

Radiant's immediate response was to suspend its lending markets on October 16 to stop further loss, engage Mandiant for incident response, and pursue tracing and recovery. Mandiant's attribution to UNC4736 / Citrine Sleet placed the attack squarely within the DPRK threat landscape. But unlike DMM Bitcoin, where a wealthy parent group could backfill the loss, Radiant was a DeFi protocol without a deep-pocketed backstop, and the ~$50M loss against a TVL that rapidly collapsed proved existential. The protocol's total value locked fell sharply in the wake of the hack as users withdrew, and over the following 18 months Radiant was unable to recover a meaningful amount of the stolen funds or raise fresh capital to rebuild; on June 1, 2026, it announced it would wind down. The most important legacy of the incident is its role as a precursor and warning that the industry largely failed to act on before Bybit. Radiant demonstrated, months before the $1.46B Bybit compromise, the exact blind-signing failure mode: compromised signer devices presenting benign transactions in the front end while malicious transactions were signed, defeating an audited Safe multisig. The remediations Radiant's experience pointed to, on-device clear-signing, independent out-of-band payload verification, and device-isolation for signers, are the same remediations the Bybit post-mortem would later make industry priorities. Radiant is, in effect, the smaller dress rehearsal that the industry under-weighted.

Industry implications and verdict

Radiant Capital's significance is disproportionate to its $50M loss because it is the clearest pre-Bybit demonstration that the multisig trust boundary extends to every signer's device and to the fidelity of what that device displays. The lessons, now canonical post-Bybit, were available here first. First, blind-signing must be eliminated for high-value operations: signers must verify the true, decoded transaction effects on a trusted device, not the front-end UI, which can be manipulated by malware. Second, signer-device integrity is a first-class control: dedicated, isolated, hardened machines for signing; no signing from a daily-driver device exposed to email, Telegram and arbitrary files; and independent verification of the transaction hash on a separate clean machine before approval. Third, the DPRK fake-contractor/recruiter social-engineering vector remains the dominant initial-access method against crypto engineering teams, and resisting it requires cultural and technical defenses against unsolicited files and code. Fourth, multisig independence is only real if signers do not share a common compromise vector; compromising the threshold via a single malware campaign collapses the security model. The verdict is that Radiant belongs in the post-mortem record as the pivotal 2024 blind-signing case, the protocol-level rehearsal for Bybit, and a cautionary tale that a $50M signer-device compromise, left as an industry-wide warning unheeded, was followed four months later by the largest theft in history exploiting the identical class of failure.

Recovery

Radiant suspended lending markets on October 16, 2024 and engaged Mandiant, but recovered no meaningful portion of the ~$50M. Lacking a deep-pocketed backstop, with TVL collapsed and unable to raise fresh capital over 18 months, Radiant announced on June 1, 2026 that it would wind down. The attackers' proceeds were largely laundered.

Key lessons

Eliminate blind-signing: verify the true decoded transaction effects on a trusted device, never the front-end UI, which malware can manipulate
Signer-device integrity is a first-class control: use dedicated, isolated, hardened machines and verify the transaction hash on a separate clean machine
Multisig independence is illusory if signers share a common compromise vector (same malware campaign, same software, same daily-driver device)
The DPRK fake-contractor/recruiter malware lure is the dominant initial-access vector against crypto engineering teams; defend culturally and technically

Frequently asked questions

What happened in the Radiant Capital Multisig/Device Compromise?

On Oct 16, 2024, Radiant Capital was drained of ~$50M across Arbitrum and BNB Chain after DPRK-aligned attackers (UNC4736 / Citrine Sleet, per Mandiant) compromised multiple signers' devices with malware (delivered via a Telegram message impersonating a former contractor). During a routine Safe multisig operation the malware showed benign transactions in the front end while the devices signed malicious ones, the exact blind-signing failure Bybit would magnify four months later. Radiant suspended markets, recovered little, and announced a wind-down in June 2026. The pivotal pre-Bybit blind-signing case study.

How much was lost?

Approximately ~$50M was lost on 2024-10-16.

How did the attack work?

Malware on multiple signers' devices that displayed benign transactions in the front end while malicious transactions were signed in the background

Who was responsible?

UNC4736 / Citrine Sleet (DPRK-aligned, assessed sub-cluster of Lazarus) per Mandiant

Were the funds recovered?