AI Agents Generate $4.6 Million in Smart Contract Vulnerabilities, Reveals Anthropic Study

Recent findings from a study conducted by Anthropic, a leading artificial intelligence firm, in collaboration with the Machine Learning Alignment & Theory Scholars (MATS), have unveiled alarming insights into the capabilities of AI agents in exploiting smart contracts. The research indicates that these AI models have autonomously generated vulnerabilities in smart contracts that could potentially lead to losses amounting to $4.6 million. This revelation raises significant concerns about the security of blockchain technologies and the rapid evolution of AI’s role in cybersecurity.

Understanding Smart Contracts and Their Vulnerabilities

Smart contracts are self-executing contracts with the terms of the agreement directly written into code. They operate on blockchain platforms, allowing for automated and trustless transactions. However, like any software, smart contracts are susceptible to bugs and vulnerabilities. These weaknesses can be exploited by malicious actors, leading to significant financial losses.

In 2026, the landscape of smart contract vulnerabilities has become increasingly complex, with AI technologies playing a pivotal role in both the development and exploitation of these contracts. The latest research indicates that the cost of exploiting these vulnerabilities is decreasing, making it easier for bad actors to launch attacks.

The Role of AI in Exploiting Smart Contracts

The study from Anthropic’s red team, which simulates the actions of malicious actors to identify potential security flaws, revealed that current commercial AI models, including Claude Opus 4.5, Claude Sonnet 4.5, and OpenAI’s GPT-5, have demonstrated the ability to exploit smart contracts effectively. During testing, these models uncovered two novel zero-day vulnerabilities in 2,849 recently deployed contracts that were initially deemed secure.

The exploits generated by these AI agents were valued at approximately $3,694, with the operational cost for GPT-5’s API being $3,476. This indicates that the financial gains from exploiting these vulnerabilities could cover the costs associated with the AI’s operation.

“This demonstrates as a proof-of-concept that profitable, real-world autonomous exploitation is technically feasible, underscoring the need for proactive adoption of AI for defense,” the research team stated.

Benchmarking AI Exploitation of Smart Contracts

To further understand the capabilities of AI in this domain, the researchers developed the Smart Contracts Exploitation (SCONE) benchmark. This benchmark consists of 405 contracts that were exploited between 2020 and 2025. When tested with ten different AI models, they successfully produced exploits for 207 contracts, resulting in a simulated loss of $550.1 million.

This benchmark highlights the increasing sophistication of AI agents in identifying and exploiting vulnerabilities in smart contracts. The study also noted a significant reduction in the number of tokens required for an AI agent to develop a successful exploit, decreasing by 70.2% across four generations of Claude models.

Implications of Decreasing Exploit Costs

The implications of these findings are profound. As the cost of generating exploits continues to decline, the window of opportunity for developers to identify and patch vulnerabilities before they are exploited will shrink. Currently, the average cost to scan a smart contract for vulnerabilities stands at just $1.22. This low cost, combined with the rising capabilities of AI, suggests that malicious actors will have more opportunities to exploit vulnerabilities before they can be addressed.

The Rapid Evolution of AI Exploitation Capabilities

The study highlights a dramatic increase in the effectiveness of AI agents over a short period. In just one year, AI agents have escalated their exploitation rate from 2% to 55.88% of vulnerabilities identified in the post-March 2025 portion of the benchmark. This leap has resulted in a staggering increase in total exploit revenue, soaring from $5,000 to $4.6 million.

Such rapid advancements in AI capabilities raise critical questions about the future of cybersecurity in the blockchain space. As AI continues to evolve, the potential for autonomous exploitation of smart contracts will likely increase, posing significant risks to developers and users alike.

Pros and Cons of AI in Smart Contract Security

While the advancements in AI present numerous challenges, they also offer potential benefits for enhancing smart contract security. Here are some pros and cons:

• Pros:
  • Enhanced Detection: AI can analyze vast amounts of data quickly, identifying vulnerabilities that human auditors might miss.
  • Automated Responses: AI can automate responses to detected vulnerabilities, potentially mitigating risks in real-time.
  • Continuous Learning: AI models can improve over time, adapting to new threats and vulnerabilities as they emerge.
• Cons:
  • Increased Exploitation: As demonstrated, AI can also be used by malicious actors to exploit vulnerabilities, leading to significant financial losses.
  • Dependence on Technology: Over-reliance on AI for security may lead to complacency among developers, who might neglect manual audits.
  • Ethical Concerns: The use of AI in malicious activities raises ethical questions about accountability and responsibility.

Conclusion

The findings from Anthropic’s study underscore the urgent need for enhanced security measures in the realm of smart contracts. As AI agents become increasingly adept at identifying and exploiting vulnerabilities, developers must prioritize proactive security strategies. This includes adopting AI-driven defense mechanisms, conducting regular audits, and staying informed about the latest advancements in AI technology.

In a rapidly evolving landscape, the collaboration between AI and cybersecurity will be crucial in safeguarding the integrity of smart contracts and protecting users from potential financial losses.

Frequently Asked Questions (FAQ)

What are smart contracts?

Smart contracts are self-executing contracts with the terms directly written into code, operating on blockchain platforms to facilitate automated transactions.

How do AI agents exploit smart contracts?

AI agents can analyze smart contracts for vulnerabilities and autonomously generate exploits, leading to potential financial losses.

What is the SCONE benchmark?

The SCONE benchmark is a framework developed to assess the exploitation capabilities of AI agents on smart contracts, consisting of contracts that were exploited between 2020 and 2025.

What are the implications of decreasing exploit costs?

As the costs of exploiting vulnerabilities decrease, the time available for developers to identify and fix these vulnerabilities before exploitation shrinks, increasing the risk of financial losses.

How can developers protect against AI-driven exploits?

Developers can enhance security by adopting AI-driven defense mechanisms, conducting regular audits, and staying updated on the latest AI advancements in cybersecurity.