Optimization and stress testing are key aspects of the design and verification process for large, high-risk systems. Optimization is about improving the capabilities and performance of a system; stress testing is about uncovering its weaknesses and faults. Both require a quantitative representation of the system's behavior, and for complex, multi-physical systems, co-simulation can be a very powerful method to create such a representation. However, co-simulation frequently involves the use of black-box subsystem models, which poses challenges to traditional optimization and stress testing methods. Here, we review the state of the art in co-simulation-based optimization and stress testing, focusing especially on \emph{adaptive stress testing} in the latter case, and discuss open research questions and promising research directions. In particular, we make the case that a co-simulation is not an entirely black box even when some or all of its subsystems are; it may be possible to exploit the visible system structure, coupling variable values, and partial subsystem information. We use examples from the maritime industry to motivate and illustrate the discussion, centering on the highly contemporary design case of an autonomous ferry.