In a groundbreaking study, advanced AI agents have successfully simulated real-world hacks in blockchain technology, reproducing exploits valued at $4.6 million. This research, conducted by Anthropic, highlights the potential of AI in identifying critical vulnerabilities in smart contracts, raising significant concerns about the future of blockchain security.
The recent tests were performed using the SCONE-bench, a comprehensive dataset consisting of 405 smart contracts that were compromised between 2020 and 2025. Researchers utilized this benchmark to evaluate ten leading AI models, instructing each to analyze and exploit the identified vulnerabilities within a controlled, sandboxed environment.
The AI agents were able to generate functioning exploits for 207 contracts, which represents approximately half of the dataset. Notably, even when researchers focused on 34 contracts that were compromised after the models” knowledge cutoff in March 2025, the AI produced successful exploits for 19 of them. The simulated value of these attacks reached a staggering $4.6 million. Among the models tested, Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5 demonstrated the most effective performance.
Furthermore, researchers examined whether these AI models could identify new vulnerabilities in smart contracts that had not yet been documented. By reviewing 2,849 recently deployed contracts on the Binance Smart Chain, the AI agents, specifically Sonnet 4.5 and GPT-5, managed to unearth two previously unknown weaknesses. These vulnerabilities generated a simulated revenue of approximately $3,694, which barely exceeded the $3,476 API cost incurred during the testing process. One flaw was linked to a public function that lacked restrictions for read-only usage, enabling unauthorized manipulation of balances. The other flaw pertained to withdrawal logic that did not adequately verify fee-recipient addresses.
All experiments were conducted on local blockchain forks, ensuring that no actual user funds were put at risk. These findings indicate that AI systems can automate the process of discovering exploits at a scale that could significantly challenge existing cybersecurity measures. The rapid identification of vulnerabilities suggests that malicious actors could launch attacks shortly after contract deployment, leaving little time for manual audits.
Experts in the field have expressed growing concerns about the implications of these developments. An ex-Apple engineer, AI Nat, cautioned that autonomous AI agents now represent a substantial threat to blockchain security. They have the ability to swiftly detect vulnerabilities, execute attacks, and adapt to patches in real time. This capability transforms security protocols into a continuous process rather than a one-time review, increasing the urgency for developers to implement constant, AI-driven monitoring to counteract evolving threats.
Another engineer, Alex Havryleshko, highlighted a significant rise in AI-related risks, noting that each incremental step in AI capabilities appears to correlate with a tenfold increase in simulated exploit revenue. He pointed out that the performance of these models seems to double approximately every 1.3 months, underscoring the rapid advancement of AI in the realm of cyber-exploitation.
Meanwhile, some commentators have noted that the high costs associated with AI agents may limit their ability to scan open-source contracts. It was also mentioned that liquidity often emerges later in the process, resulting in narrow detection windows. They emphasized that addressing vulnerabilities during the development phase with AI tools represents the most effective strategy for defense, as easily exploitable targets are rapidly diminishing.
