Tobias Braun

Multimodal foundation models are increasingly deployed in settings where they generate images, interact with software tools, and produce formal outputs such as code. This thesis studies the robustness and security of such models under realistic adversarial objectives. I develop and evaluate backdoor threat models in which concept erasure for text-to-image diffusion can be circumvented, allowing adversaries to recover supposedly removed content. I further study multimodal backdoors in unified autoregressive models with shared parameters for text and image generation, where triggers can propagate across modalities and jointly manipulate visual outputs and accompanying text. I then extend these robustness questions to computer-use multimodal agents that act through tool calls, visual interfaces, and memory, including multi-step attacks that exploit benign-looking actions and manipulated intermediate rationales to evade oversight. The project is informed by my earlier work on multimodal fact-checking, where reliability depends on evidence grounding and robustness to spurious cues in tool-augmented pipelines.