AI safety is a system property, not a sum of individual defenses.

Safety for large language models (LLMs) and artificial intelligence (AI) agents remains fragmented; defenses against jailbreaking coexist with unresolved hallucination, fairness, and autonomous-control problems. Existing frameworks address these risks, but few specify how failure in one dimension changes another within a deployment. We argue that distinct safety failures can interact and amplify each other. We identify four safety challenge classes, analyze their interactions, and ground the argument in real-world incidents involving compound failures. We then audit GPT-5 and Claude-Sonnet-4.6 on clinical vignettes from the Medical Question Answering benchmark based on the United States Medical Licensing Examination (MedQA-USMLE). A single jailbreak can affect multiple safety dimensions, although patterns vary by model and attack. We propose three design principles for integrated safety architectures and identify agent safety and mechanistic interpretability as urgent open frontiers.