This project evaluates the safety, robustness, and reliability of vision-language models (VLMs) like GPT-4o and GPT-4o-mini in Out-of-Distribution (OOD) and challenging scenarios. Using the VLLM Safety Benchmark, it explores their performance on datasets such as OODCV-VQA, Counterfactual VQA, and Sketchy-VQA, highlighting their limitations in handling counterfactual reasoning, abstract sketches, and ambiguous inputs. The findings aim to identify safety concerns, quantify model capabilities, and contribute to the development of secure, reliable VLMs for real-world applications involving nuanced and complex visual contexts.