AI Exploitation of Ethereum Smart Contracts: GPT-5 and Claude Unveil Million-Dollar Vulnerabilities

Recent developments have shown that artificial intelligence (AI) can exploit vulnerabilities in Ethereum smart contracts, posing significant economic risks. A collaborative study by Anthropic and MATS Fellows utilized a new metric called the Smart CONtracts Exploitation benchmark (SCONE-bench) to evaluate the capabilities of AI models. The research revealed that these AI systems could potentially cause financial damage amounting to millions of dollars by identifying and exploiting hidden weaknesses in smart contracts.

Key Takeaways

AI models, including GPT-5 and Claude, have successfully identified vulnerabilities in Ethereum smart contracts, leading to simulated exploits worth millions.
The SCONE-bench framework measures financial damage from exploits rather than traditional metrics, highlighting the economic impact.
The study indicates that AI can autonomously exploit blockchain systems, leading to significant security risks.
Developers are urged to employ AI-driven defensive measures to protect against potential vulnerabilities.

Ethereum smart contracts exploited by AI: GPT-5 and Claude demonstrate million-dollar vulnerabilities - 1

The Rise of AI in Exploiting Smart Contracts

Artificial intelligence is increasingly capable of targeting vulnerabilities within smart contracts on Ethereum and other blockchains, prompting serious discussions about the economic risks associated with automated cyber operations. In a recent study conducted by Anthropic in collaboration with MATS Fellows, the researchers introduced the Smart CONtracts Exploitation benchmark (SCONE-bench) to analyze the effectiveness of AI models against 405 real-world smart contracts that were exploited between 2020 and 2025.

During simulated attacks against contracts that had been compromised after March 2025, advanced AI agents—Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5—collectively uncovered exploits estimated to yield a staggering $4.6 million, establishing a plausible minimum for the potential financial harm that AI could inflict. When the scope was expanded to include 2,849 recently launched contracts without previously identified vulnerabilities, GPT-5 and Sonnet 4.5 managed to expose two novel zero-day vulnerabilities, leading to simulated profits of nearly $3,700.

SCONE-bench: Measuring Financial Risk

Unlike traditional cybersecurity benchmarks that assess effectiveness by detection rates or subjective scores, SCONE-bench uniquely quantifies AI exploits based on potential monetary losses, offering a clearer picture of risk. Smart contracts are especially suitable for this examination since vulnerabilities can result in direct monetary theft, allowing simulations to effectively estimate possible losses.

Across the 405 contracts assessed in SCONE-bench, 10 different AI models succeeded in exploiting 207 contracts, with total simulated losses reaching $550.1 million. Even when factoring in possible data contamination, leading-edge models consistently demonstrated their ability to exploit contracts that were beyond their training knowledge.

Specific Instances of AI Exploitation

One notable instance involved a vulnerability within a token calculator function on an Ethereum-compatible contract, which had mistakenly been left writable. An AI agent exploited this gap, repeatedly invoking the function to artificially inflate its token balance, resulting in simulated earnings of $2,500. Under optimal liquidity conditions, this could have resulted in profits of up to $19,000. Fortunately, white-hat intervention later secured the recovered assets.

The findings emphasize the fact that AI agents are nearing human-level proficiency in tasks such as analyzing control flows, evaluating boundaries, and exploiting software vulnerabilities—skills that are applicable to both blockchain technology and traditional software environments.

The research underscores the rapid advancement of AI's cyber capabilities, evolving from simple network breaches to the autonomous exploitation of blockchain systems. SCONE-bench serves as a vital resource, enabling smart contract developers to conduct stress tests on their systems prior to deployment.

Researchers opine that this study provides a compelling proof-of-concept for the potential for profitable, real-world autonomous exploitation, spotlighting the urgent necessity for proactive AI-based defensive strategies to safeguard financial systems and digital assets.