A February 2026 study by AI security firm Cecuro found that a purpose-built AI security agent successfully detected 92% of real-world DeFi exploits tested in its research.
The study analyzed 90 smart contracts that had been exploited between late 2024 and early 2026, representing approximately $228 million in total losses.
Outperforming General AI Models
According to the findings, the specialized AI system significantly outperformed general-purpose models such as GPT-5.1, which identified only 34% of the same vulnerabilities.
The vulnerabilities correctly flagged by Cecuro’s system accounted for $96.8 million of the total exploited value, indicating that the tool was able to identify many of the most economically damaging flaws.
Researchers attributed the strong performance not to raw computational power alone, but to domain-specific security layers and methodologies tailored specifically for smart contract analysis.
Open Research, Restricted Tool
Cecuro has made its research data and evaluation framework publicly available on GitHub as open-source material. However, the full detection tool remains restricted.
The company stated that limiting access to the production-grade system is intended to prevent malicious actors from using advanced vulnerability detection capabilities to identify and exploit weaknesses before they are patched.
Defensive AI vs. Offensive AI
The study arrives amid growing concerns that “offensive” AI tools, used by hackers to scan code and identify attack vectors, are advancing rapidly.
As DeFi protocols continue to manage billions of dollars in on-chain value, the race between automated attack tools and automated defense systems is intensifying.
The findings suggest that specialized, security-focused AI agents may play an increasingly central role in protecting smart contract infrastructure as threat sophistication continues to evolve.
