Worked example · 2026-09

Nostra — a token worth $550K in total was accepted as collateral for $3.5M of loans after its price moved 8,000× in minutes — Nostra Finance / Starknet — 2026-09-17

Loss

~$3.5M borrowed against inflated NSTR collateral by a single account, across ETH, STRK, USDC, USDT, WBTC and DAI. NSTR's entire market cap was ~$550–590K at the time. ~$1.92–1.93M was bridged to Ethereum (including 234.57 ETH and 1M DAI); ~2.2M STRK left via NEAR Intents; ~$1.55M remained on Starknet at last report. Nostra paused supply, borrow, withdrawal and liquidation.

OAK Techniques observed

OAK-T9.001 (Oracle Price Manipulation — primary, confirmed class. The NSTR price the protocol consumed rose from ~$0.006 to ~$49.5 (~8,000×) in minutes, driven through a thin NSTR/SolvBTC pool. See techniques/T9.001-oracle-price-manipulation.md). Which feed Nostra used for NSTR is not established: Nostra integrated Chainlink in May 2026 for eight majors but not NSTR, and reporting has not clarified whether the NSTR feed came from Pragma, an aggregator, or elsewhere. One analysis states a fake NSTR/SolvBTC pool hijacked GeckoTerminal's pool-selection logic; OAK records that as reported, not confirmed, pending Nostra's post-mortem.

Attribution

pseudonymous.

Key teaching point

Borrowable value against a collateral asset must be capped by what that asset can actually be sold for, not by what its price feed says. No oracle design survives a 6× gap between loan size and total market cap: even a perfectly honest price for NSTR could not have been liquidated into $3.5M. A supply cap sized to exit liquidity on the protocol's own governance token would have bounded this regardless of feed choice. And a feed that can move 8,000× in minutes without a deviation breaker refusing to act on it is not a price, it is an input.