Blockchain has suffered a series of high-impact financial attacks in recent years, underscoring the urgent need for robust security mechanisms. As it becomes increasingly central to native financial infrastructures such as cryptocurrencies and decentralized finance (DeFi), ensuring the financial security of blockchain ecosystems has emerged as a critical challenge. While a growing body of research has focused on smart contract vulnerabilities, relatively limited attention has been paid to the detection of anomalous behaviors at the account level–despite their central role in on-chain financial threats. Existing anomaly detection approaches predominantly rely on static rules or pre-defined patterns, which are inadequate for modeling the dynamic and covert behaviors of malicious accounts. To address this gap, we propose AccountCatcher, an anomaly detection framework based on graph embedding and graph attention network(GAT). By modeling account interactions and capturing abnormal transactional patterns, our approach effectively identifies high-risk accounts. Experiments on real-world Ethereum data demonstrate that AccountCatcher attains 94.4% accuracy and a 93.9% F1 score, surpassing the strongest baseline, Trans2Vec, by 7.1% and 6.8%, respectively, providing a more reliable safeguard for detecting financial anomalies in blockchain ecosystems.