Andreessen Horowitz (a16z) has issued a stark warning that artificial intelligence agents are now capable of autonomously identifying and reproducing vulnerabilities in decentralized finance (DeFi) protocols. In a research paper published by a16z crypto, the venture firm revealed that AI agents achieved a nearly 70% success rate in replicating simple price manipulation attacks when provided with structured knowledge about common exploit patterns. However, the agents still struggle with complex multi-step attacks.
The current security model for DeFi, which relies on point-in-time audits, is no longer sufficient, a16z argues. AI agents can continuously test exploit paths without waiting for scheduled reviews, widening the gap between attackers and defenders. The firm advocates for a shift toward security based on formal specifications and invariants, proving what a protocol is allowed to do rather than reacting after an attack.
Composability between DeFi protocols amplifies the risk, according to a16z. A vulnerability in a single contract can become systemic when exploited across bridges, lending pools, or oracles. AI agents evaluate whether a system's assumptions break down at machine speed, and they do not distinguish between core and peripheral failures.
The warning comes amid a surge in DeFi losses. Security firm GoPlus Security reported that four separate smart contracts on Ethereum mainnet were exploited within 48 hours ending April 29, with combined losses exceeding $1.5 million. The firm described the current pace of AI-assisted attacks as a 'countdown-by-the-second era.'
Among the incidents, a ZetaChain-related exploit saw approximately $333,868 drained across nine transactions on Ethereum, Arbitrum, Base, and BSC. ZetaChain confirmed that no user funds were lost—the three affected wallets belonged to the team—and that the vulnerability had been previously reported via its bug bounty program but was dismissed. The protocol has paused cross-chain transactions and is deploying a patch.
Other exploits include an onchain aggregator contract losing roughly $983,000 due to missing access controls, an unauthorized third-party vault tied to TradingProtocol losing about $398,000, a BCB contract losing $39,800 from a reentrancy vulnerability, and a QNT asset contract losing $124,900 from an arbitrary call vulnerability.
Compounding concerns, Anthropic recently announced a new AI model called 'Claude Mythos Preview,' capable of autonomously finding and writing working exploits for zero-day vulnerabilities across major operating systems and web browsers. The company confirmed that the same improvements that make the model effective at patching vulnerabilities also enable it to exploit them.
Cryptopolitan reports that DeFi losses in April alone have reached record levels, surpassing the combined stats for the first three months of the year. The escalating crisis is setting the stage for an AI arms race between hackers and developers.
