The series of attacks on the Ethereum mainnet that led to over $1.5 million in losses has been exacerbated by new research that shows that artificial intelligence (AI) agents can now autonomously discover and exploit vulnerabilities in decentralized finance protocols.
Security firm GoPlus Security reported that four separate contracts were exploited in just 48 hours ending April 29. The firm warned that hackers armed with AI are becoming more precise and faster than ever.
And DeFi smart contract developers have nowhere to turn to except AI to tackle the problems that AI itself started.
Can AI really hack DeFi by itself?
a16z crypto tested an off-the-shelf AI coding agent against 20 past price manipulation incidents on Ethereum and found that when given just a contract address and basic tools, the AI succeeded in exploiting the vulnerability only 10% of the time.Â
However, when researchers gave the agent access to structured knowledge about common attack patterns like vault donation exploits and automated market maker (AMM) pool manipulation, the success rate jumped to 70%.
The researchers noted that while the AI is very good at finding bugs, it sometimes struggles with complex, multi-step attacks. One agent even tried to “escape” its test environment by extracting a secret key to look at future block data.
Anthropic recently announced a new AI model called “Claude Mythos Preview.” The company stated that this model can autonomously find and write working exploits for zero-day vulnerabilities across major operating systems and web browsers.Â
Before Mythos Preview, older models had a “near-0% success rate” at writing exploits. The company also confirmed that the same improvements that make the model good at patching vulnerabilities also make it good at exploiting them.Â
When given access to Etherscan’s transaction API, the agent found actual past attack transactions and reverse-engineered them to write its own exploit code.Â
How much was lost in the ZetaChain hack?
GoPlus Security flagged four separate smart contract exploits on Ethereum mainnet within a 48-hour window ending April 29. The combined losses exceeded $1.5 million. The firm has described the current pace of AI-assisted attacks as a “countdown-by-the-second era.”Â
In one of the week’s larger incidents, approximately $333,868 was drained across nine transactions on four chains, including Ethereum, Arbitrum, Base, and BSC. ZetaChain’s official post-mortem report says that no user funds were lost; the three affected wallets belonged to the ZetaChain team.Â
The attacker took advantage of a feature in the GatewayEVM contract using “arbitrary calls.” The gateway lacked a strict blocklist, allowing the hacker to instruct it to transfer token allowances that had been set by the team wallets.
The hacker funded wallets through Tornado Cash three days before the attack while mimicking a victim’s wallet.Â
ZetaChain admitted that the vulnerability had been reported earlier through its bug bounty program, but the initial reports were dismissed. The protocol has since paused cross-chain transactions and is rolling out a patch to disable the risky code.Â
Other Ethereum exploits identified by GoPlus Security over the past 48 hours include an onchain aggregator contract that lost roughly $983,000 due to missing access controls; an unauthorized third-party vault tied to TradingProtocol that lost roughly $398,000 also due to missing permission checks; a BCB contract that lost roughly $39,800 from a reentrancy vulnerability; and a QNT asset contract that lost roughly $124,900 from an arbitrary call vulnerability.Â
Cryptopolitan reports that DeFi losses in April alone have reached record levels, surpassing the combined stats for the first three months of the year.
With mounting losses in recent cases, it is setting up an epic showdown where hackers and developers are fighting AI with AI. With Anthropic’s Mythos and others now entering the conversation, it is looking like AI is arming hackers and developers won’t have any choice but to use AI to defend themselvesÂ
