AI Models Achieve Human-Level Exploit Capability in Smart Contracts, Simulating $550M in Theft

In a groundbreaking study, AI research company Anthropic has demonstrated that advanced AI systems can effectively hack smart contracts, replicating historical exploits and discovering new vulnerabilities, raising fresh concerns about DeFi security.

The research, conducted with MATS and Anthropic Fellows, used the SCONE-bench benchmark comprising 405 smart contracts that were actually hacked between 2020 and 2025. When ten leading AI models, including Claude Opus 4.5, Claude Sonnet 4.5, GPT-5, Llama 3, and DeepSeek V3, were tested in a simulated environment, they managed to exploit 207 out of 405 contracts, with the simulated value of stolen funds reaching approximately $550 million.

To ensure the models were not simply recalling past incidents, the team focused on 34 contracts exploited after March 1, 2025. On this set, Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5 still produced working exploits on 19 contracts, worth a combined $4.6 million in simulated value, with Opus 4.5 alone accounting for about $4.5 million.

Anthropic further tested the agents on a zero-day dataset of 2,849 recently deployed Binance Smart Chain contracts with no known vulnerabilities. Both Sonnet 4.5 and GPT-5 found two zero-day bugs, generating attacks worth $3,694, with GPT-5 achieving this at an API cost of about $3,476. All tests were conducted on forked blockchains and local simulators, with no real funds at risk.

The study found that potential exploit revenue on the 2025 problems roughly doubled every 1.3 months, while the token cost of generating exploits fell sharply across model generations. For instance, across four generations of Claude models, token costs decreased by 70.2%.

David Schwed, COO of SovereignAI, commented: "AI is already being used in security tools, and bad actors will use the same technology to identify vulnerabilities. This can all be done now 24/7, against all projects." However, he added that with proper controls and testing, most exploits are avoidable, and good actors have the same access to AI for defense.

Anthropic emphasized that the capabilities enabling smart contract exploitation also apply to traditional software, and urged developers to adopt automated tools in their security workflows to keep pace with offensive advancements.