AI Agents Exploit Smart Contracts for $1.22, Redefining DeFi Security

Anthropic's Frontier Red Team has demonstrated that AI agents can autonomously exploit smart contracts, fundamentally altering the landscape of DeFi security. These agents, trained to emulate professional attackers, learned to fork blockchains, write exploit scripts, and drain liquidity pools in simulated environments. A recent study revealed that frontier models like Claude Opus 4.5 and GPT-5 reconstructed 19 out of 34 real smart contract exploits that occurred after March 2025, extracting $4.6 million in simulated value, without prior knowledge of the vulnerabilities.

The economic implications are terrifyingly viable. GPT-5 scanned 2,849 BNB Chain ERC-20 contracts for approximately $1.22 per contract, uncovering two novel zero-day vulnerabilities. The average cost per vulnerable contract identified was $1,738, yielding a net profit of around $109 per exploit at current capabilities, a figure projected to double every 1.3 months. One example zero-day involved a rewards token with a public “calculator” function missing a “view” modifier, allowing users to inflate their balances and drain liquidity, with an estimated $2,500 to $19,000 extractable value.

These agents operate in containerized environments, analyzing contract logic, identifying state transitions, and constructing multi-step transaction sequences. They successfully chained flash loans, manipulated oracle prices, and exploited reentrancy across multiple contracts, demonstrating a sophisticated understanding of Solidity semantics and DeFi composability. The key change is automation; an AI agent can produce a working proof of concept in under 60 minutes, a task that would take human auditors hours. Across 405 real exploits, models generated working exploits for 207 contracts, totaling $550 million in simulated theft.

To counter this evolving threat, Anthropic proposes three main strategies: integrating AI-driven fuzzing into CI/CD pipelines using tools like SCONE-bench for continuous testing; drastically shortening patch and response cycles to achieve sub-hour detection and remediation; and recognizing that this automation race extends beyond DeFi to broader cybersecurity domains. The study underscores that security is no longer a one-time audit but an continuous adversarial engagement, where defenders must adopt AI capabilities to outpace automated attackers.

(Source: https://cryptoslate.com/anthropic-ai-smart-contract-exploits/)