XNews

Anthropic Red Team Research: AI Attacks on Smart Contracts

Anthropic Red Team Research: AI Attacks on Smart Contracts

Research by the Anthropic Red Team and MATS Fellows shows that AI can automatically discover and exploit vulnerabilities in smart contracts for profit. In a test set containing 405 real vulnerabilities, AI successfully breached 207 contracts. If this occurred on a real blockchain, it would correspond to approximately $550 million in stolen funds. For 34 new vulnerabilities that emerged after March 2025, AI successfully exploited 19 (55.8%), earning about $4.6 million, with Claude Opus 4.5 contributing around $4.5 million. During a scan of 2,849 brand-new contracts, GPT-5 and Claude Sonnet 4.5 each discovered 2 previously unknown zero-day vulnerabilities, simulating profits of approximately $3,694. The cost for GPT-5 to complete this scan was about $3,476, nearing break-even. Research statistics indicate that from 2024 to 2025, AI attack profits double every 1.3 months, token consumption decreases by an average of 23.4% every two months, and overall attack efficiency improves by 70.2%.

Key insights

Research Objective: Quantifying the Economic Value of AI Attacks

This study, conducted by the Anthropic Red Team and MATS Fellows, focuses on smart contracts to quantify the economic attack capabilities of AI. The reasons for selecting smart contracts include the public availability of contract source code, the direct association of vulnerabilities with economic losses, and the ability to replicate in a simulated environment without causing real damage. The study aims to assess whether AI can independently discover and exploit contract vulnerabilities for profit.

High Success Rate and Significant Potential Losses from Known Vulnerability Tests

The test set containing 405 real vulnerabilities showed that AI successfully breached 207 contracts. If this occurred on a real blockchain, the corresponding estimated stolen amount would be around $550 million, indicating that in scenarios with known vulnerabilities, AI's attack efficiency and economic impact are significant.

AI Achieves Attacks and Near-Break-Even Cost Structures Against New and Zero-Day Vulnerabilities

In the study, AI successfully exploited 19 out of 34 vulnerabilities that emerged after March 2025, earning approximately $4.6 million (with Claude Opus 4.5 contributing about $4.5 million). During the scanning of 2,849 brand-new contracts, GPT-5 and Claude Sonnet 4.5 each discovered 2 unknown vulnerabilities, simulating profits of around $3,694, while the scanning cost for GPT-5 was about $3,476, indicating that attacks are nearing economic break-even.

Trends Indicating Rapid Progress in Attack Efficiency and Cost Metrics

Research statistics indicate that from 2024 to 2025, AI attack profits double every 1.3 months, token consumption decreases by an average of 23.4% every two months, and overall attack efficiency improves by 70.2%, suggesting a rapid evolution in attack capabilities and cost efficiency.

Source posts