Microarchitectural vulnerabilities such as Spectre and Rowhammer continue to endanger systems, with new variants repeatedly evading existing defenses. In this talk, I will emphasize the need for robust offensive testing to proactively uncover such vulnerabilities in future hardware. First, I will introduce AMuLeT (ASPLOS’25) <https://gururaj-s.github.io/assets/pdf/ASPLOS25_FuTanenbaum.pdf>, an automated framework that detects speculative execution leaks in CPUs at design time. By extending model-based relational testing to simulators like gem5, AMuLeT enables the systematic testing of Spectre defenses. In the first large-scale evaluation of four widely used mitigations, AMuLeT uncovered 3 known and 6 previously unknown vulnerabilities, including the first reported insecurity in SpecLFB (SEC’24). Next, I will present GPUHammer (USENIX Security ’25) <https://gpuhammer.com/>, the first Rowhammer attacks on discrete GPUs. This attack can degrade ML model accuracy from 80% to 0.1%, with a single bit flip in GPU memory, revealing a new class of threats to AI safety and reliability. Finally, I will present QPRAC (HPCA’25) <https://arxiv.org/abs/2501.18861>, a low-overhead Rowhammer defense for future DRAM based on the PRAC framework. Together, these efforts show how rigorous offensive testing can enable robust defenses.