Applications of neural networks in automating control tasks are one the most popular topics in software-intensive companies these days. With the integration of deep reinforcement learning (DRL) methods, classic algorithms were outperformed in various software-related tasks such as playing complex games, self-driving cars, etc. One of the interesting applications of reinforcement learning is its application as an automated stock trading software system. However, these algorithms are prone to manipulations by competing trading systems, similar to player vs. player settings. But, verifying the quality of DRL-based commercial trading software is challenging for a third party since they are protected behind secure international exchange APIs such as NASDAQ. This makes typical gradient-based adversarial sample generation (as a quality assessment tool for Deep Neural Networks), like FGSM, obsolete. In this research, we aim to demonstrate a “grey-box” approach for attacking a deep reinforcement learning trading agent is possible by trading in the same stock market. We assume we don’t have access to the trading agents’ source code, policy architecture, Deep Neural Network weights, or training algorithm. Our only viable data would be the current state of the market and the decision of the trading agent or its chosen action in that given state. The adversary agent is provided with the combination of inputs to the trading agent’s Deep Neural Network policy and the trade decision made by the agent. Our proposed adversary the agent uses a hybrid Deep Neural Network as its policy consisting of convolutional layers and fully-connected layers. The adversary policy proposed in our research was able to change the softmax output of the baseline by 44.66%, ensemble method by 28.7%, and the automated trading software developed by our industrial partner Intelius AI by 21.2% and reducing their profits by 87.94%, 74.46%, and 66.82%, respectively.