The field of foundation models is moving towards a greater understanding of their vulnerabilities, particularly in regards to backdoor attacks and dataset poisoning. Recent research has highlighted the potential risks of these attacks, which can compromise the security and reliability of models. The development of new attack methods and benchmarks has shed light on the weaknesses of current models, including their sensitivity to textual instructions and vulnerability to backdoor triggers. Notably, researchers have found that even minimally poisoned datasets can lead to highly vulnerable models, underscoring the need for more research into the robustness of foundation models. Some noteworthy papers in this area include: GFM-BA, a novel backdoor attack model that effectively addresses the challenges of launching backdoor attacks against Graph Foundation Models. BackdoorVLM, a comprehensive benchmark for evaluating backdoor attacks on vision-language models, which has revealed significant vulnerabilities in current models.