In 92% of DeFi exploits AI security review flags underlying problem (www.coindesk.com)

A recent study by AI security firm Cecuro has demonstrated that a specialized AI agent can detect vulnerabilities in 92% of real-world decentralized finance (DeFi) exploits, outperforming a general-purpose GPT-5.1 coding agent that only achieved 34% detection. The research evaluated 90 compromised smart contracts, revealing that the targeted AI flagged vulnerabilities associated with $96.8 million in exploit value, while the baseline agent uncovered only $7.5 million. The significant difference in performance stemmed from the application of domain-specific security methodologies and DeFi-focused heuristics layered on top of the same underlying AI model. This finding raises critical concerns for the AI/ML community, particularly as the prevalence of AI-enabled cybercrime, including sophisticated smart contract exploits, accelerates. Research indicates that exploit capabilities using AI are doubling approximately every month, with low-cost attack attempts further lowering barriers for malicious actors. The growing reliance on general-purpose AI and one-off audits for security, as highlighted by Cecuro, may leave many DeFi projects vulnerable to intricate attacks that have evaded prior professional scrutiny. By releasing the benchmark dataset and evaluation framework on GitHub, Cecuro aims to spur innovation in defensive tools, albeit cautiously withholding its full security agent out of concern for potential misuse in offensive attacks.