# A Faulty Pragma Price Feed Liquidated $3 Million of Collateral on Starknet in Two Minutes

Vesu says 47 borrowing positions were closed between 04:08 and 04:10 UTC on 4 September, and that its own contracts contained no bug.

- Published: 2026-09-06T06:01:25.324Z
- Canonical: https://polylog.news/crypto/2026-09-06/a-faulty-pragma-price-feed-liquidated-3-million-of-collatera
- Publisher: Polylog (Crypto desk)
- Section: crypto
- Sources: [crypto.news](https://crypto.news/vesu-oracle-incident-triggers-3m-in-liquidations/), [Cryptonomist](https://en.cryptonomist.ch/2026/09/05/vesu-oracle-incident-liquidations/)

Vesu, a lending protocol on the Starknet layer-2 network, said an upstream price source operated by the oracle provider Pragma delivered incorrect data for about two minutes on 4 September. During that window the wrong prices reached several Vesu liquidity pools and [made 47 borrowing positions appear eligible for liquidation](https://crypto.news/vesu-oracle-incident-triggers-3m-in-liquidations/). Automated liquidator bots took roughly $3 million of collateral before the feed returned to the correct value.

Vesu's assessment is that no contract behaved incorrectly. The liquidation engine executed exactly as written, on inputs that were wrong. The protocol says it is [coordinating with Pragma, StarkWare, the Starknet Foundation and the curators of the affected pools](https://en.cryptonomist.ch/2026/09/05/vesu-oracle-incident-liquidations/) to recover the seized assets. It has not said how much is recoverable, how the recovery would work, or whether any liquidator has agreed to return anything. Liquidators who acted through open, permissionless functions have no obligation to give the collateral back.

The pattern repeats. Aave suffered a comparable failure in March, when a stale parameter triggered an estimated $26 million to $27 million of unintended liquidations of wrapped staked ether, after which that protocol reviewed its oracle update rates and fallback logic. In both cases the audited on-chain code was fine and the loss came through a privileged input the protocol trusts by design.

Starknet holds $403.7 million of total value secured, up 7.5% over seven days according to L2Beat, so a $3 million forced liquidation is a meaningful share of the collateral on that network. The incident also raises a governance question. Recovery depends on social coordination among a foundation, an oracle firm and pool curators, which is a reminder of how much of a nominally trustless system runs on identifiable parties agreeing to act.

## What this means

Lenders and borrowers on any chain are exposed to whoever writes prices into the contract, and that party is usually outside the scope of the audits and bug bounties the protocol advertises. The people who lost collateral here made no error in position management. Two outcomes decide the cost of this failure class: either lending protocols add safeguards such as multiple independent feeds, deviation limits and liquidation delays, which reduce capital efficiency, or they keep single-provider feeds and depositors continue to absorb occasional two-minute losses with no recourse.

## What to watch

- Whether any of the $3 million is returned, which would show that informal recovery works on small chains where participants are identifiable, or that it does not.
- Whether Pragma publishes a root-cause report naming the upstream source that failed, since oracle providers rarely disclose which venue produced bad data.
- Whether other Starknet lending markets add deviation caps or delay windows before liquidation, the concrete change that would prevent a two-minute glitch from closing positions.