The rapid expansion of decentralized finance (DeFi) has introduced unprecedented opportunities for financial inclusion, yet the irreversible and permissionless nature of blockchain networks has simultaneously made them a primary target for sophisticated fraud, including rug pulls, phishing drains, and layering schemes. Traditional manual compliance and rule-based systems are increasingly inadequate against the speed of on-chain attackers. This study proposes an advanced AI-driven detection framework that utilizes temporal graph modeling and behavioral data fusion to identify malicious activities in real-time. Using a substantial dataset of over 1.25 million unique addresses and 18.4 million transactions across the Ethereum, BSC, and Solana blockchains, the research evaluates the efficacy of Graph Neural Networks (GNNs) and unsupervised autoencoders. The methodology emphasizes a "ground truth" labeling process and chronological data splitting to prevent data leakage. Results indicate that the hybrid systemincorporating on-chain transaction traces, smart contract bytecode, and off-chain social signalsachieves an F1-score of 0.81 with a remarkably low false-positive rate of 0.008. Furthermore, the study demonstrates operational efficiency with an inference time of 18ms, making it viable for high-throughput networks. The findings suggest that while AI significantly enhances blockchain security, ongoing challenges in "explainability" and adversarial obfuscation require a "human-inthe-loop" approach to maintain long-term ecosystem trust.