Anthropic Discovers AI Can Exploit Smart Contracts, Identifying $4.6M in Vulnerabilities

In an alarming revelation, Anthropic has illustrated that advanced AI systems can detect vulnerabilities in blockchain applications, potentially leading to significant financial exploits. This raises critical questions about the security of decentralized finance (DeFi) platforms.

Key Takeaways

Anthropic's research highlights AI's capability to find weaknesses in smart contracts, potentially leading to significant financial losses.
The study employed a benchmark called SCONE-bench with 405 compromised smart contracts from the years 2020 to 2025.
AI agents were able to exploit over half of the reviewed contracts, resulting in simulated thefts totaling around $550.1 million.
Newer AI models, including Claude Opus 4.5 and GPT-5, identified $4.6 million in vulnerabilities from newly created contracts.
All AI testing was conducted in controlled environments, with no real funds at risk, highlighting the technical possibilities without impacting live systems.
The study emphasized that while AI can exploit vulnerabilities, it also holds potential for auditing and fixing smart contracts.

Anthropic

AI Identifies Vulnerabilities in Blockchain Applications

Anthropic’s recent study, conducted alongside MATS and Anthropic Fellows, demonstrated the efficacy of AI agents in locating weaknesses in blockchain-based applications. These findings raise important concerns regarding the overall security of DeFi platforms.

The research employed a benchmark known as SCONE-bench (Smart CONtracts Exploitation) that utilized a dataset of 405 smart contracts compromised between 2020 and 2025. Upon testing 10 advanced models in a simulated environment, the AI agents successfully exploited over half of the smart contracts, leading to simulated thefts amounting to approximately $550.1 million.

To ensure the findings were relevant, the researchers focused on 34 contracts that had been exploited after March 1, 2025, which was the latest knowledge cutoff for these AI systems.

New Exploit Targets Discovered

In this more focused analysis, the AI models, including Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5, successfully executed attacks on 19 contracts, amounting to a simulated total of $4.6 million. Notably, Opus 4.5 accounted for an impressive $4.5 million of the identified vulnerabilities.

Further testing evaluated whether these AI agents could identify new issues instead of revisiting past weaknesses. On October 3, 2025, Sonnet 4.5 and GPT-5 were deployed against 2,849 newly launched Binance Smart Chain contracts, all of which had no previously identified vulnerabilities.

Test Simulation

Both AI agents detected two zero-day vulnerabilities, generating attack scenarios valued at $3,694, with GPT-5's operational cost at approximately $3,476.

Safeguards Implemented in Testing

It's important to note that the testing occurred on simulated blockchains and forked networks rather than live environments, ensuring that no actual funds were at risk. Anthropic's goal was to assess what is technically achievable today while avoiding interference with functioning systems.

Smart contracts provide a viable testing ground, as they handle significant value and operate entirely on-chain. When smart contract code fails, attackers can often directly withdraw assets. Researchers can replicate these actions and translate stolen assets into dollar amounts using historical data, providing concrete estimates of potential damages from AI-driven exploits.

The SCONE-bench framework evaluates success based on financial outcomes, rather than simple pass/fail criteria. AI agents are assigned code, context, and tools within a sandbox environment, with a successful run requiring the agent to accumulate at least 0.1 additional ETH or BNB, preventing minor errors from being recorded as significant successes.

Economics of Attacks and Evolving Threats

The study revealed that the potential financial gain from exploits in 2025 has roughly doubled every 1.3 months, while the costs associated with generating viable exploits have significantly decreased across evolving AI models. This implies that attackers can execute more effective attacks within the same computational budget as technology progresses.

Although the research primarily focuses on DeFi, Anthropic indicates that the methodologies used can extend to conventional software, covering public APIs and lesser-known internal services.

Anthropic emphasizes to the cryptocurrency development community that while these AI tools can facilitate exploitation, they also have the potential to aid in auditing and correcting vulnerabilities before deployment.