An exploit on Starknet lending protocol Nostra on September 17 drained roughly $3.5 million, and the post-mortem points at a problem that sits under most on-chain lending: a token's oracle price and the price you can actually sell it for are two different numbers. Oracle provider Pragma published an assessment on September 18 that flagged 6 of its 22 mainnet price feeds as critical risk, naming its own feeds as part of the failure chain.
The mechanics of the Nostra drain
The attacker used NSTR, Nostra's own token, as collateral. By pushing the NSTR oracle price up, a single account was able to borrow around $3.5 million in assets against collateral that was not worth anywhere near that amount once anyone tried to sell it. Nostra paused lending, borrowing, withdrawals, and liquidations the same day. Final loss figures and any recovery were still unconfirmed at the time of Pragma's report.
This is a familiar attack shape. An oracle reports a value, the lending market treats that value as gospel, and someone borrows real assets against a number that only holds up on paper.
The gap Pragma measured
Pragma's September 18 liquidity report is the part worth reading closely. Instead of asking "is the price feed accurate," it asked a harder question: if a liquidator had to sell this collateral right now, at what price would the sale actually clear?
The answer was ugly for several tokens. Comparing a $10,000 sale against a $10 baseline quote, the reports showed the effective price falling sharply:
- LORDS: about 22% deterioration
- BROTHER: about 20% deterioration
- EKUBO: about 17% deterioration
- NSTR: about 15% deterioration
A feed can be perfectly accurate about the last trade and still be dangerous, because the last trade was tiny. Nobody liquidates a bad loan by selling $10 at a time.
Which feeds got flagged
Pragma tagged 6 of its 22 mainnet feeds as critical risk: BROTHER, DAI, DOG, EKUBO, LORDS, and NSTR. Nine more feeds received a high-risk rating. The presence of DAI on that critical list is notable, since a widely held stablecoin ending up in the same bucket as thin governance tokens says the rating is about on-chain sell depth on Starknet specifically, not about the asset's reputation elsewhere.
For a lending market, a critical-risk feed is not a warning to ignore. It is a live invitation to the exact borrow-against-inflated-collateral move that hit Nostra.
The lesson for anyone lending on-chain
Depositing tokens into a lending protocol hands custody of those assets to a smart contract and ties your safety to that protocol's oracle design. That is a different risk than holding assets in your own wallet, where no external price feed decides whether you get liquidated. If you supply liquidity or lend on any chain, the questions that matter are not just the headline APY. They are: which oracle secures this market, does it price collateral off deep liquidity or a thin pool, and can a small manipulation move the number enough to unlock a large loan.
Pragma deserves some credit for publishing feed-level risk ratings and quantifying the liquidity gap rather than quietly patching. Transparency about which feeds are fragile is more useful to depositors than a vague "we take security seriously" note. The uncomfortable takeaway is that this gap exists across many protocols and chains, not just this one incident. Most lending markets still price collateral off feeds that assume you can sell at the quoted price. Nostra is the example that got tested.
Overview
A manipulated NSTR oracle price let an attacker borrow about $3.5 million from Nostra on Starknet on September 17. Pragma's September 18 assessment flagged 6 of its 22 mainnet feeds (BROTHER, DAI, DOG, EKUBO, LORDS, NSTR) as critical risk and showed several tokens losing 15-22% of their quoted value on a $10,000 sale versus a $10 baseline. The core issue is structural: oracle prices do not guarantee that collateral can be liquidated at those prices, and that gap sits under most on-chain lending, not just this protocol.
