Energy theft causes large economic losses to utility companies around the world. In recent years, energy theft detection approaches based on machine learning (ML) techniques, especially neural networks, become popular in the research literature and achieve state-of-the-art detection performance. However, in this work, we demonstrate that the well-perform ML models for energy theft detection are highly vulnerable to adversarial attacks. In particular, we design an adversarial measurement generation algorithm that enables the attacker to report extremely low power consumption measurements to the utilities while bypassing the ML energy theft detection. We evaluate our approach with three kinds of neural networks based on a real-world smart meter dataset. The evaluation result demonstrates that our approach can significantly decrease the ML models' detection accuracy, even for black-box attackers.