logo

ChainThink

Stay ahead, master crypto insights

Anthropic: AI Agent Discovers $4.6 Million Vulnerability in Real-World Smart Contract

Anthropic: AI Agent Discovers $4.6 Million Vulnerability in Real-World Smart Contract

2025-12-02 10:00

View Original
Anthropic released a report stating that its researchers tested the Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5 models on a self-built SCONE-bench benchmark (featuring 405 real-world exploited smart contracts from 2020 to 2025). After the knowledge cutoff date (March 2025), the three models collectively identified exploitable vulnerabilities worth approximately $4.6 million in affected contracts. Additionally, in simulated testing of 2,849 recently deployed contracts with no known vulnerabilities, Sonnet 4.5 and GPT-5 each discovered 2 novel zero-day vulnerabilities, collectively posing potential losses of $3,694, with GPT-5's API cost amounting to $3,476.
#Vulnerability Detection

Disclaimer: Contains third-party opinions, does not constitute financial advice

Recommended Reading
Anthropic announces $100 million investment into the Claude Partner Network
Anthropic announces $100 million investment into the Claude Partner Network
GoPlus Launches SafuSkill, Building a Security-First AI Agent Skills Marketplace
GoPlus Launches SafuSkill, Building a Security-First AI Agent Skills Marketplace
Anthropic Makes Cover of TIME Magazine: Dubbed "the Most Disruptive Company in the World," Valued at $380 Billion, Surpassing Goldman Sachs and Coca-Cola
Anthropic Makes Cover of TIME Magazine: Dubbed "the Most Disruptive Company in the World," Valued at $380 Billion, Surpassing Goldman Sachs and Coca-Cola
Anthropic Launches "Anthropic Institute": Co-founder Jack Clark Appointed as Head of Public Benefit, Focusing on AI's Impact on Employment, Law, and Economics
Anthropic Launches "Anthropic Institute": Co-founder Jack Clark Appointed as Head of Public Benefit, Focusing on AI's Impact on Employment, Law, and Economics
Anthropic to Open Permanent Office in Washington and Triple Public Policy Team Size
Anthropic to Open Permanent Office in Washington and Triple Public Policy Team Size
The White House Plans Executive Order to Fully Eliminate Anthropic, with Jeff Dean and Other Competitors' Employees Uniting in Public Support
The White House Plans Executive Order to Fully Eliminate Anthropic, with Jeff Dean and Other Competitors' Employees Uniting in Public Support
Claude announces the launch of Code Review feature, enabling AI agents to automatically conduct code reviews
Claude announces the launch of Code Review feature, enabling AI agents to automatically conduct code reviews