Ajna's oracle-free design loses $775,000 to its own liquidation arithmetic

Ajna does not remove the price function when it removes the oracle; it moves that function inside the contracts, into the pool's own bucket and take accounting. In Ajna, lenders deposit into buckets at prices they pick, and the contracts read that state to decide when a loan is liquidatable, with whoever starts a liquidation posting a bond so that calling a healthy loan carries a cost ^[7][8]. The design deletes a third-party feed, which is real, but it also promotes the pool's own take-and-bucket arithmetic into the price function, and Ajna's published audit history already contained findings on that surface: take computations during liquidation, and bucket state accounted for incorrectly, both since marked resolved ^[10]. Nethermind's summary of this class of attack, that they force the contract to calculate a distorted price and exploit it before the transaction ends ^[11], describes something that needs no external feed to be present.

The size is what argues against a one-off. Roughly $775,000 in ETH against the $206,000 of Ajna V2 TVL that DefiLlama showed at the time is 3.8 times the recorded deposit base ^[1][15][17], and the loss was spread: syrupUSDC gave up about $173,700, or 22.4% of the total ^[9][14], which leaves $601,300 across the six other named pools, near $100,200 each if it were even ^[2][12]. Repetition across pools reads as a procedure rather than a lucky transaction. (The at-the-time figures also sit oddly with each other: about $418,000 of active loans against $206,000 of TVL is 203% ^[15][18], a useful reminder that headline TVL is not the collateral pile a reader assumes it is.)

Defimon says it spotted a prepared attack more than an hour before the first exploit transaction and posted the warning in Ajna's Discord, and the protocol was still unsecured when the first one landed ^[21]. An hour is a long time if there is a lever and no time at all if there is not, and the white paper's promise of a system that requires no governance or external price feeds to function ^[6] is the reason to suspect the second. What is left afterwards is a loan book that has gone from about $418,000 to $30,200, down 92.8% ^[16][20], now standing at 6.7% of a $450,000 TVL ^[19], which is borrowers using one dollar in fifteen of the shelf.

This is probably wrong in one direction, or rather, the more interesting version of wrong. The MixBytes premise, that a significant share of DeFi attacks come out of oracle price manipulation, configuration errors and access control problems ^[3], is not refuted by a single incident, and the oracle-having comparison in the same account is Moonwell, where roughly $7 million pushed an illiquid token eight-fold, borrowed close to $10 million of real assets and left with about $6 million, an 86% return on capital deployed ^[4][13], which is an order of magnitude dearer than what happened here, so relocating an attack surface and worsening it remain separate claims. What would show me wrong is a post-mortem that names one arithmetic bug in the take path, patches it in a commit, and demonstrates no analogous invariant sitting unguarded elsewhere. What would confirm it is the next oracle-free lender losing money through accounting its contracts trusted because they computed it themselves.