The intersection of artificial intelligence (AI) and cryptocurrencies is expanding significantly.
For example, in October, CriptoNoticias reported on a project that lets AI agents trade Bitcoin (BTC) and cryptocurrencies.
In this case, a new experiment published on December 1 by Anthropic, the company that created the Claude model, shows that AI agents can do more than analyze data.
Human researchers have revealed that AI algorithm was able to exploit vulnerabilities in smart contracts scale.
By testing 405 real-world contracts deployed on networks such as Ethereum, BNB Chain, and Base between 2020 and 2025. generated model script 207 of them are equipped with functional attack equipment.51.1% indicated “success”.
By executing these attacks in a controlled environment that reproduces network conditions called scone bench,,, The simulated losses amounted to approximately $550 million.
The findings highlight threats to decentralized platforms (DeFi) and smart contracts, Build in automated defenses.
Details of experiments using AI and virtual currency networks
Experimental methods incorporate AI models such as Claude Opus 4.5 and GPT-5. instructed to generate exploit (the code that exploits the vulnerability) in an isolated container (Docker) and uses a time limit of 60 minutes per attempt.
In addition to testing previously hacked contracts, we also included new contracts with no known flaws to look for vulnerabilities. “Zero Day” (unknown).
The following graph shows the rapid improvement in the effectiveness of state-of-the-art models. trace Total simulated profit This is what each main model was able to generate (on a logarithmic scale) by exploiting all vulnerabilities in the test suites used to evaluate the performance of different AI models.
This image shows an exponential trend. Recent models such as GPT-5 and Claude Opus 4.5 have achieved hundreds of millions of dollars in simulation profits, far exceeding previous models such as GPT-4o.
Further experiments confirmed this potential “income”. Doubles approximately every 0.8 monthshighlighting that advances in offensive capabilities are accelerating.
The second graph, on the other hand, details the performance on a more difficult subset: vulnerabilities discovered in 2025.
Here, the metric called “Pass@N” measures the success of generating multiple pass attempts. exploit (N attempts) By contract. This explains how the total simulated revenue steadily increases as more attempts (from Pass@1 to Pass@8) are allowed, reaching $4.6 million.
The second graph confirms the following: Claude Opus 4.5 was the most effective model in this controlled environmentachieve the largest portion of their profits.
Finally, this study shows that the probability of exploitation is not correlated with code complexity, but is. The amount of funds held by the contract. This model focuses on contracts with higher lock values and tends to detect attacks more easily.