AI Agents Find Smart Contract Exploits, Uncovering Millions in Simulated Vulnerabilities
Anthropic research shows AI agents like Claude Opus 4.5 and GPT-5 can autonomously exploit smart contract vulnerabilities, collectively identifying $4.6 million in simulated exploits.
Intelligence analysis by Gemini 2.5 Flash
A new benchmark, SCONE-bench, was developed to evaluate AI agents' ability to exploit smart contracts, quantifying the financial impact of AI cyber capabilities. The study found that frontier AI models could exploit real-world vulnerabilities and even uncover novel zero-day exploits in simulated environments, demonstrating the technical feasibility of autonomous, profitable exploitation.
Imagine a super-smart computer program that's really good at finding tiny mistakes in other computer programs, especially ones that handle money on the internet, like a digital piggy bank. This program found ways to trick these digital piggy banks in a pretend world, making it seem like it stole millions of dollars. It even found new mistakes no one knew about! This shows that while these smart programs can be tricky, we also need to use them to protect our digital money.
Analysis
Anthropic's recent research, conducted as part of a MATS and Anthropic Fellows project, delves into the economic implications of AI models' growing cyber capabilities, specifically focusing on smart contract exploitation. The study introduces a novel benchmark, SCONE-bench, designed to quantify the financial consequences of AI-driven vulnerabilities, moving beyond traditional success rates to measure impact in monetary terms. This approach offers a more tangible assessment of risk for policymakers, engineers, and the public.
SCONE-bench
SCONE-bench is a groundbreaking benchmark comprising 405 smart contracts that were actually exploited between 2020 and 2025 across Ethereum, Binance Smart Chain, and Base blockchains. This benchmark allows for direct measurement of the dollar value of exploits by running them in simulated environments, making smart contracts an ideal testing ground due to their public source code and direct financial implications. The evaluation framework uses Docker containers for sandboxed execution, ensuring reproducible results and offering plug-and-play support for auditing contracts before deployment. This direct quantification of losses, using on-chain assets, provides a more accurate and less speculative measure of vulnerability impact compared to bug bounty programs or theoretical models.
$4.6 million
Using SCONE-bench, frontier AI models, including Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5, collectively developed exploits worth $4.6 million in simulated funds from contracts exploited after their knowledge cutoffs. This figure establishes a concrete lower bound for the potential economic harm these AI capabilities could enable. Beyond retrospective analysis, the agents were also tested against 2,849 recently deployed contracts without known vulnerabilities, where they uncovered two novel zero-day vulnerabilities and produced exploits worth $3,694. Notably, GPT-5 achieved this at an API cost of $3,476, demonstrating the potential for profitable, real-world autonomous exploitation. The research observed that exploit revenue from stolen simulated funds roughly doubled every 1.3 months over the last year, indicating a rapid acceleration of these capabilities.
Balancer
The article provides a concrete example of a real-world smart contract exploit, referencing the Balancer incident in November 2025. In this event, an attacker exploited a rounding direction issue to steal over $120 million from users' funds. This example underscores the severe financial consequences that smart contract vulnerabilities can have. The researchers argue that assessing AI agents on smart contract exploitations provides a lower bound on the economic impact of their broader cyber capabilities because smart contract and traditional software exploits draw on a similar set of core skills, such as control-flow reasoning, boundary analysis, and programming fluency. The findings from SCONE-bench thus serve as a critical warning and a call to action for proactive adoption of AI for defensive cybersecurity measures.
Key points
- AI agents, including Claude Opus 4.5 and GPT-5, can autonomously exploit smart contract vulnerabilities.
- A new benchmark, SCONE-bench, quantifies the economic impact of AI cyber capabilities by measuring simulated stolen funds.
- AI models collectively identified $4.6 million in simulated exploits from previously exploited contracts.
- Agents uncovered novel zero-day vulnerabilities in recently deployed contracts, demonstrating real-world exploitation feasibility.
- The research emphasizes the urgent need for proactive adoption of AI for defensive cybersecurity in the blockchain space.
The research highlights the potential for AI to be a powerful tool for defense, suggesting that AI agents could be used to audit smart contracts for vulnerabilities before deployment, thereby preventing real-world exploits and enhancing blockchain security. Proactive adoption of AI for defensive purposes could significantly mitigate the risks posed by malicious AI capabilities.
The study demonstrates the technical feasibility of profitable, autonomous exploitation of smart contracts by AI agents, posing a significant risk of economic harm in the real world. The rapid acceleration of AI cyber capabilities could lead to increasingly sophisticated and frequent attacks, potentially overwhelming current human-led defense mechanisms.
