The Expanding Audit Surface: A Survey of Safety Auditing Across Models, Agents, and Deployed Systems

Safety and security are often conflated in the literature on large language model (LLM) systems, although spontaneous system failures and adversarial compromise require different evidence and controls. This survey asks how safety auditing changes as an LLM is embedded in a tool-using agent, a multi-agent collective, and a domain-specific deployed system. It introduces a four-scope framework that separates model, agent, collective, and deployment auditing and relates audit evidence to governance decisions. We conduct a structured narrative review of roughly 370 works, classifying them by unit of analysis, audit timing, instrument, evidence source, and deployment domain. Six comparison tables and three figures synthesize the literature. Adversarial security work is used only to define the survey boundary or assess audit instruments. Auditing methods are most mature at model scope and become less standardized as system complexity increases. Agent and collective benchmarks remain young and inconsistently defined; tool and skill ecosystem auditing is especially recent; and deployment evidence is concentrated in a small set of domains. Automated-judge reliability and benchmark validity recur as constraints across all four scopes. Reliable oversight requires audit units, evidence, and governance commitments to expand with the system being deployed. The four-scope framework identifies where current methods transfer across levels and where continuous, domain-specific assurance remains missing.