Generative artificial intelligence (AI) offers a powerful avenue for peptide design, yet this process remains challenging due to the vast sequence space, complex structure-activity relationships, and the need to balance antimicrobial potency with low toxicity. Traditional approaches often rely on trial-and-error screening and fail to efficiently navigate the immense diversity of potential sequences. Here, we introduce AMP-Diffusion, a novel latent diffusion model fine-tuned on antimicrobial peptide (AMP) sequences using embeddings from protein language models. By systematically exploring sequence space, AMP-Diffusion enables the rapid discovery of promising antibiotic candidates. We generated 50,000 candidate sequences, which were subsequently filtered and ranked using our APEX predictor model. From these, 46 top candidates were synthesized and experimentally validated. The resulting AMP-Diffusion peptides demonstrated broad-spectrum antibacterial activity, targeting clinically relevant pathogens-including multidrug-resistant strains-while exhibiting low cytotoxicity in human cell assays. Mechanistic studies revealed bacterial killing via membrane permeabilization and depolarization, and the peptides showed favorable physicochemical profiles. In preclinical mouse models of infection, lead peptides effectively reduced bacterial burdens, displaying efficacy comparable to polymyxin B and levofloxacin, with no detectable adverse effects. This study highlights the potential of AMP-Diffusion as a robust generative platform for designing novel antibiotics and bioactive peptides, offering a promising strategy to address the escalating challenge of antimicrobial resistance.