This chapter explores machine unlearning (MU) as a principled mechanism for enhancing the safety, reliability, and compliance of foundation models such as large language models (LLMs) and diffusion models (DMs). We examine three key dimensions where MU plays a critical role. First, hazardous knowledge removal targets the erasure of unsafe capabilities, such as harmful visual concepts in DMs, thereby mitigating risks while preserving general utility. Second, spurious correlation removal addresses shortcut learning in safety fine-tuning, where models become overly sensitive to superficial cues, producing “safety mirages” that cause both adversarial jailbreaks and excessive conservatism. Third, copyright and privacy protection highlights MU’s role in removing copyrighted material and personally identifiable information, balancing legal compliance with model usability. Across these applications, empirical studies demonstrate that MU complements traditional safety alignment methods by directly erasing risky knowledge rather than merely steering model behavior. These applications position MU as an effective tool for aligning large-scale generative models with safety, ethical, and regulatory requirements.