Anthropic trained its AI models to detect vulnerabilities in smart contracts and reported they identified flaws enabling approximately $4.6 mln in potential theft.

Testing and results

The company tested models on actual buggy contracts originating from Ethereum and BNB Chain to assess detection performance in realistic conditions.

Models discovered roughly half of known vulnerabilities and additionally flagged nineteen distinct issues in previously unseen contracts during the evaluation period.

Anthropic estimated that the vulnerabilities identified correspond to attack vectors that could have enabled thefts totaling about $4.6 mln, according to the report.

Model performance

Among the models, Claude Opus 4.5 demonstrated the highest effectiveness in locating exploitable code paths and other critical errors.

• Scope: real buggy contracts from Ethereum and BNB Chain.
• Detection rate: approximately fifty percent of known vulnerabilities.
• New findings: 19 issues in contracts previously unknown to the models.

Implications

Detection rate reached approximately fifty percent, leaving a substantial portion of vulnerabilities undetected and requiring additional mitigation steps by developers.

The stated objective is to produce developer-facing tools that surface vulnerabilities before deployment, thereby reducing the risk of on-chain exploits.

Anthropic intends to integrate model outputs into pre-deployment analysis workflows to help developers remediate high-risk issues before publishing contracts.