The attacker's route through Avici reportedly ran SubmitSignatures on the authorization program, then AddCollateralAdmin on the collateral program, then WithdrawCollateralAsset [6], a sequence that asks a program to believe its own collateral ledger rather than to defeat a signature, and DefiLlama filed it as a withdrawal logic flaw in a Rust-based protocol rather than as anything more glamorous [4]. A day later, Defimon described Ajna's loss as liquidation accounting manipulation [9]. The chain was different and the dollar amount was different, but the trust that failed was the same kind in both cases.
Which is why the CoinGecko security figures are the load-bearing arithmetic here, or rather the ratio inside them: 147 of the 245-plus incidents logged between January 2025 and July 2026 hit audited protocols, and those 147 carried 88.44% of the $3.63 billion stolen [13][14], about $3.21 billion [9] from 60% of the events [8], with the report noting that most attacks went at infrastructure, third parties, governance or human error rather than at code inside the audit's scope [14]. The pool meant to absorb that has thinned: active on-chain cover fell from $163.2 million to $130.2 million, a decline of 20.2% [15][6], with five of nine on-chain insurers inactive or pivoted by August 2026 [15]. Against $3.63 billion of losses, $130.2 million is 3.6% [7].
So the refund is a balance-sheet decision, not an insurance recovery. Divide $500,859.22 by 1,685 affected users and the promise costs about $297.25 each [1], which is cheap enough that the interesting question is what it prices for everyone else, because a firm that loses a hundred times as much cannot buy the same sentence. The attacker's wallet, holding roughly 10,005 SOL worth about $1.07 million plus $11,600 in stablecoins [7], comes to about 2.16 times the logged loss [2], consistent either with earlier balances in that wallet or with the initial estimates of $600,000 to more than $1 million that Avici has not confirmed [5].
Ajna is the more brutal comparison. Its roughly $775,000 loss is 3.14 times the $246,880 still locked in the protocol [9][12][4], total value locked was already down 71.3% over the prior 30 days [12], and Defimon says it flagged a prepared attack more than an hour before the first exploit transaction and warned the team in Discord without response [10]. The depositors best served were the ones who left before any of that.
This is probably wrong, but I read Avici's pledge as the cheapest reputational asset on the market at these dollar amounts, and the token tape supports the cheapness: AVICI still sits more than 96% below its November 2025 peak of $7.61 [8], so the equity holders paid for the incident and the cardholders are being made whole out of what is left. How this reads later depends on what happens next. If the 1,685 refunds settle on-chain at the stated figure, the floor is real and the next issuer answers for it. If they settle quietly, partially, or only at the lower of the competing loss estimates [5], the pledge was a press cycle. And if another program running Rain's flawed contract version turns up with its own hole [3], the $500,859 stops being the incident and starts being the first invoice.
