← Breach analysis by the Dual Vulnerabilities framework

Radiant Capital flash-loan exploit (Unattributed, 2024)

2024 · Canada · K - Financial and insurance activities · Flash-loan-based protocol exploit via smart-contract rounding error

Analysed 2026-11-06 by Hermes

Open this case in the tool →

What happened

Flash-loan-based protocol exploit via smart-contract rounding error

Dual vulnerability type

Expertise to Execution

Evidence coverage: 1/6 stages, weight 8/60

Decisive stage: entry (clearest evidence) · Alternative: Strategy to Execution

Why this classification

(1) Radiant Capital is a cross-chain DeFi lending protocol built as a fork of the Aave/Compound code lineage and holding user deposits across chains; the specialist expectation for such a protocol is that the known failure classes of its fork lineage - above all the empty-market / first-depositor rounding vulnerability, whose root cause predates this incident - are closed before a new market goes live, and the project presented exactly that doctrine publicly: over $1.5M spent on code audits, contests and white-hat bounties, professional auditors and a contracted real-time monitoring stack. (2) In practice, when the new native USDC market on Arbitrum was activated on 2 January 2024 it was empty, and an attacker with a pre-deployed contract struck within about 5 seconds: repeated flash loans inflated the liquidityIndex of the zero-deposit reserve and rounding in favor of the user at burn let the attacker withdraw far more collateral than posted, leaving 1900 WETH (~$4.5M, 1.3% of TVL) of bad debt; the seed deposit that would have closed the vector was not part of the launch process. (3) This is a knowing-doing gap inside the material column, not generic weak security: the remedy was standard, publicly known specialist practice (PeckShield: the root cause was not new), it required no leadership decision or new budget - Radiant itself shipped the fix within days as a deployment-process change - yet the new-market deployment was executed without it; the classification rests on the entry stage, the only stage with a documented gap, while containment (DAO emergency pause), detection (Hypernative flag, same-evening acknowledgment), recovery (markets resumed 4 January, bad debt repaid under RFP-27) and response/disclosure are documented as effective and produce no compensating gap. The entry vector is fully known and load-bearing here. (4) RM29 (Expertise to Execution lens): the rounding/precision handling in the protocol's own deployed contract code (rayDiv rounding and rounding in favor of the user when burning rTokens) enabled the drain, and the empty-market countermeasure that is standard secure-development practice for this code lineage was not applied to the new market's code and launch sequence (causal role: Enabling). RM20 (Expertise to Execution lens): activating a new market was a release that altered the security posture - a fresh, empty, borrowable reserve - and the change went live without a dedicated new-market review or validation gate (seed deposit, zero collateral factors), steps the post-mortem introduced only afterwards, so the environment changed faster than it was hardened against an attacker watching governance (causal role: Enabling). RM12 (Expertise to Execution lens): assurance did not show real exposure: the team pointed to over $1.5M of audits, contests and bounties by specialist firms, yet the known empty-market class in the audited fork lineage was not caught or rated critical, leaving the team without reliable evidence of the market's safety at launch (causal role: Obscuring).

Evidence by stage

Risk markers

Sources

Notes

Description matches the stated victim; the row's narrative (flash loan, $4.5M ETH, rounding error, temporary halt) is confirmed by the victim's own post-mortem and same-day reporting. Amounts: 1900 WETH / ~$4.5M per the post-mortem, PeckShield and Beosin (bad debt later stated as 1,910 ETH total: 1,190 repaid + ~720 remaining). The case file's third source_url (beincrypto.de, '50 Millionen') is titled after the separate October 2024 incident, not this one - a provenance caveat for the dataset row, not a description mismatch. Country taken from eurepoc_extra.receiver_country (Canada); Radiant operates as a DAO with no verified headquarters, and the related row 2047 carries Hong Kong for the same victim - the jurisdiction was not independently verifiable, so the dataset field is kept as-is and not guessed further. Related row 2047 (case file and its existing result opened): a later, separate incident (16 October 2024, multisig signer compromise by Citrine Sleet, ~$50M); the weakness class is not comparable to the 2024-01 rounding exploit, so RM21 is not coded on either row - noted only as a prior/later incident at the same victim within 2024. Actor genuinely unattributed: Radiant offered a $100,000 bounty via Arkham and involved Chainalysis and law enforcement, with no public attribution reported; this does not lower the DV assessment, which is victim-side. Link check: medium.com returns HTTP 403 to some automated fetches; the post-mortem page was opened and read during this research. Link check: cointelegraph.com was unreachable at check time (timeout); the page was opened and read during research. Link check: theblock.co was unreachable at check time (timeout); the page was opened and read during research. Link check: ct.com was unreachable at check time (timeout); the page was opened and read during research. Unattributed actor per initiator_name 'Not attributed'; no attribution found during research.