Worked on the Azure/PyRIT repository to enhance the reliability of gradient-based computations in the GCG Attack module. Addressed a critical bug by standardizing the handling of infinity values, replacing np.infty with np.inf in scenarios involving non-ASCII tokens. This Python-based fix improved numerical stability and prevented erroneous gradient updates, reducing the risk of NaN or infinite values in downstream machine learning analyses. Leveraging expertise in cybersecurity and machine learning, the update ensured more consistent and reliable defense and attack simulations within PyRIT. The work focused on code correctness and maintainability, contributing to the robustness of gradient computation workflows.