Swiss Cheese Model for AI Safety: A Taxonomy and Reference Architecture for Multi-Layered Guardrails of Foundation Model Based Agents

Multiple safety layers are necessary, not just one: The paper applies the "Swiss Cheese Model" to AI safety, meaning that no single guardrail is perfect—each has gaps. Therefore, effective AI safety requires multiple overlapping safeguards at different stages of how AI agents operate, so that if one fails, others can catch problems.

AI agents need safety checks across their entire pipeline: The research identifies that guardrails must operate at multiple points where foundation model-based agents make decisions, including when they set goals, create plans, select tools, and take actions. A single checkpoint at the end is insufficient given the complexity of modern AI systems.

Safety must be built into design from the start: The paper emphasizes "AI-safety-by-design from a software architecture perspective," suggesting that safety considerations should be integrated into how AI systems are engineered from the beginning, rather than added as an afterthought to existing systems.

Current guardrail design faces real challenges: The autonomous and unpredictable nature of foundation model-based agents makes it difficult to create effective safeguards. The paper's systematic review reveals that designing runtime guardrails is complicated by the involvement of multiple pipeline stages and agent components that interact in non-deterministic ways.

A structured framework can guide safer AI development: The paper provides both a taxonomy of safety attributes and a reference architecture that researchers and practitioners can use as concrete guidance when building AI systems, helping standardize how safety considerations are addressed across different applications.