Anthropic's AI agents autonomously generate $4.6M in smart contract exploits, raising alarm

AI Agents Demonstrate Unprecedented Capability in Smart Contract Exploits

Anthropic has revealed findings that mark a significant shift in blockchain security threats. The company's research shows that AI agents using Claude Opus 4.5, Claude Sonnet 4.5, and OpenAI's GPT-5 autonomously generated smart contract exploits worth $4.6 million when tested on contracts exploited after March 2025—the training data cut-off date for these frontier AI models 1

1

. The tests demonstrated that AI agents matched the performance of skilled human attackers in more than half of the smart contract exploits recorded on major blockchains over the last five years 2

2

. This capability surge represents a dramatic acceleration, with exploit revenue from stolen simulated funds roughly doubling every 1.3 months over the past year 1

1

.

SCONE-Bench Reveals Alarming Success Rates Across Ethereum and Beyond

To measure these capabilities systematically, Anthropic introduced SCONE-bench, a Smart CONtracts Exploitation benchmark consisting of 405 smart contracts on three Ethereum-compatible blockchains: Ethereum, Binance Smart Chain, and Base 1

1

. When ten frontier AI models were evaluated against this dataset, they collectively produced working attacks against 207 contracts, totaling $550 million in simulated stolen funds 2

2

. The dataset derives from the DefiHackLabs repository of contracts successfully exploited between 2020 and 2025, providing a realistic testing ground for assessing how effectively AI agents can find and exploit smart contract vulnerabilities 1

1

. Anthropic's strongest model, Claude Opus 4.5, exploited 17 of the 34 post-March 2025 vulnerabilities and accounted for $4.5 million of the total simulated value 2

2

.

Zero-Day Vulnerabilities Discovered at Profitable Rates

The cybersecurity risks from AI extend beyond known vulnerabilities. When researchers tested Claude Sonnet 4.5 and GPT-5 against 2,849 recently deployed contracts with no publicly disclosed vulnerabilities, both AI agents uncovered two novel zero-day vulnerabilities and produced exploits worth $3,694 1

1

. The automated attacks proved economically viable, with GPT-5 achieving its results at an API cost of $3,476—meaning the exploits nearly covered their own operational expenses 3

3

. The average cost per agent run came to just $1.22, while the average cost per vulnerable contract identified was $1,738, and the average revenue per exploit reached $1,847, yielding an average net profit of $109 1

1

. One discovered flaw involved a token contract with a public calculator function lacking a view modifier, allowing the agent to repeatedly alter internal state variables and sell inflated balances on decentralized exchanges, generating about $2,500 in simulated value 2

2

.

Falling Exploitation Costs Accelerate Threat Timeline

The economics of AI-driven attacks are shifting rapidly in favor of malicious actors. Analyzing four generations of Claude models, researchers found the median number of tokens required to produce a successful exploit declined by 70.2% 3

3

. This reduction in computational overhead directly translates to falling exploitation costs, making these automated attacks increasingly financially appealing. David Schwed, COO of SovereignAI, noted that the model-driven attacks would be straightforward to scale because many vulnerabilities are already publicly disclosed through Common Vulnerabilities and Exposures or audit reports 2

2

. He emphasized that bad actors can now operate 24/7 against all projects, targeting even those with smaller total value locked because the attacks are agentic and require minimal human oversight 2

2

.

Blockchain Security Faces an Asymmetric Challenge

Anthropic's findings demonstrate that in just one year, AI agents progressed from exploiting 2% of vulnerabilities in the post-March 2025 portion of their benchmark to 55.88%—a leap from $5,000 to $4.6 million in total exploit revenue 3

3

. Most smart contract exploits from 2025 could have been executed autonomously by current AI agents . The research indicates that attackers can now probe any contract interacting with valuable assets, including authentication libraries, logging tools, or long-neglected API endpoints 4

4

. Schwed characterized the issues as "business logic flaws," explaining that AI systems can identify these weaknesses when given structure and context about how a smart contract should function 2

2

.

AI-Driven Defense Strategies Emerge as Essential Countermeasure

Anthropic's research team concluded with a call for proactive adoption of AI for defense, arguing that the same capabilities enabling exploitation can strengthen security 1

1

. The company plans to open-source its SCONE-bench dataset to help developers benchmark and harden smart contracts . Security experts acknowledge that AI is already being used in Application Security Posture Management tools like Wiz Code and Apiiro, as well as in standard Static Application Security Testing and Dynamic Application Security Testing scanners 2

2

. Schwed pushed back against doom-and-gloom narratives, stating that with proper controls, rigorous internal testing, real-time monitoring, and circuit breakers, most vulnerabilities are avoidable, since good actors have access to the same agents as bad actors 2

2

. With falling costs and rising capabilities, the window between vulnerable contract deployment and exploitation will continue to shrink, leaving developers less time to detect and patch vulnerabilities 3

3

.