Moonwell lost $1.78 million after a pricing error hit its cbETH Core Market on Base. The protocol said cbETH was valued at roughly $1.12 instead of its real market price near $2,200, letting liquidators seize large amounts of collateral while repaying very little debt.
The issue appeared after governance proposal MIP-X43 was implemented. During the integration of Chainlink’s Oracle Extractable Value (OEV) wrapper contracts, the system failed to calculate cbETH’s USD price correctly. Rather than multiplying the cbETH/ETH exchange rate by the ETH/USD feed, it treated the cbETH/ETH rate as if it were already a dollar-denominated price. That mistake sharply undervalued collateral and opened the door to abusive liquidations.
Faulty pricing let liquidators extract outsized collateral
According to Moonwell, an attacker could repay about $1 of debt and pull out thousands of dollars worth of cbETH, WETH, USDC, and other assets. Within hours, the protocol cut the cbETH borrow cap to 0.01 to halt new borrowing and contain the damage. The response was quick, but liquidations that had already gone through left users with material losses.
Moonwell said part of the affected code had been written by Anthropic’s Claude Opus 4.6 model. That shifted attention to the review process around AI-assisted development. The problem was not limited to model output; developers and auditors also failed to catch the oracle formula error before the code reached production.
Review and audit layers failed to stop the bug
Smart contract auditor Pashov said there is still a person behind the AI checking the final work, and possibly an auditor as well, so blaming the neural network alone is inaccurate. Even so, the incident has added to concerns around “vibe coding.” Blockchain security firm SlowMist described the case as an oracle formula vulnerability and tied it to weak human oversight that allowed flawed code into a live system.
The case also fits into a wider pattern of coding risks tied to AI tools in crypto. The source cited a study that found 69 vulnerabilities across 15 applications built with AI coding tools such as Cursor, Claude Code, and Codes. It also referenced Anthropic research published in December 2025, which said Claude Opus 4.5 could independently simulate exploits worth $4.6 million.
Parallels drawn with Bithumb’s ghost-value error
The report also pointed to Bithumb’s February 6 mistake in South Korea, where a wrong-unit assignment in a rewards promotion created more than $40 billion in ghost value. The two incidents were different in mechanics, but they point to the same pressure point: a small error in pricing logic or unit handling can produce outsized damage once it reaches live financial systems.