Swiss Cheese Model for AI Safety: Multi-Layered Runtime Guardrail Taxonomy
Shamsujjoha, Lu, Zhao, Zhu (CSIRO Data61) · v2025-01 (v3) · first published 2024-08-05
Scope
A SLR-derived taxonomy of multi-layered runtime guardrails for foundation-model-based agents. Organized along three architectural dimensions (pipelines, artifacts, actions) plus quality attributes that guardrails should satisfy. Evoked as a "Swiss Cheese" defense-in-depth reference architecture for AI safety by design.
Why it's ingested
The Swiss Cheese Model paper gives SAICA-KG's Tool ontology an external architectural anchor for defense-in-depth reasoning: Tool nodes can be mapped to (pipeline stage x artifact x action) triples and evaluated against the 14 quality attributes. Strong crosswalks: SC-PIPE-01/02/03 aligns directly with SAICA-KG's TemporalPhase partition; SC-ART-* overlaps with LocusOfControl; SC-ACT-* overlaps with ControlParadigm (Block -> prevention, Flag -> detection, Modify/Retry -> correction, Fall-back/Isolate -> recovery). This makes the Swiss Cheese paper a first-class validation anchor for SAICA-KG's axis choices.
Categories & crosswalks
Each row lists one external category and any SAICA-KG FailureModes that cross-walk into it.
SC-QA-01AccuracyMitigates hallucinations, misinformation, and disinformation to ensure factually correct outputs.
SAICA-KG crosswalks:SC-QA-02EfficiencyPrevents resource-intensive tasks and endless loops; manages cost through dynamic resource management.
SAICA-KG crosswalks:SC-QA-03PrivacyProtects against sensitive-data leakage via direct responses or statistical inference.
SAICA-KG crosswalks:SC-QA-04SecurityDefends against malicious activities and adversarial attacks; preserves data confidentiality and integrity.
SAICA-KG crosswalks:SC-QA-05SafetyPrevents harmful or misleading outputs, especially in high-stakes domains.
SAICA-KG crosswalks:SC-QA-06FairnessAddresses bias and discrimination from training data, algorithms, or deployment.
SC-QA-07ComplianceEnsures adherence to legal, regulatory, and industry standards via continuous monitoring.
SC-QA-08GeneralizabilityFunctions effectively across multiple layers and diverse scenarios without per-case configuration.
SC-QA-09CustomizabilityTailored protection for specific requirements at different operational layers.
SC-QA-10AdaptabilityAdjusts dynamically to varying conditions and emerging threats without manual reconfiguration.
SC-QA-11TraceabilityDetailed logs of decision paths, inputs, outputs, and data sources for auditing.
SC-QA-12PortabilityGuardrail transfers across FM-based agents with minimal adjustment.
SC-QA-13InteroperabilityFunctions seamlessly across heterogeneous agents, technologies, and interfaces.
SC-QA-14InterpretabilityProvides clarity on how a guardrail operates, enabling user understanding and trust.
SC-PIPE-01Prompt StageIncoming user queries analyzed for harmful content and safety alignment before processing.
SC-PIPE-02Intermediate Results StageOutputs at workflow stages verified for accuracy and safety before final generation.
SC-PIPE-03Final Results StageFinal agent outputs checked for goal alignment and governance compliance.
SC-ART-01GoalsDeclared user or task objectives.
SC-ART-02ContextConversation and session context.
SC-ART-03MemoryLong-term or episodic agent memory.
SC-ART-04ReasoningChain-of-thought and intermediate reasoning traces.
SC-ART-05PlansMulti-step action plans produced by the agent.
SC-ART-06WorkflowsEnd-to-end orchestrated task flows.
SC-ART-07ToolsExternal callable tools and APIs.
SC-ART-08Knowledge BasesRetrieval corpora and structured knowledge stores.
SC-ART-09Other AgentsPeer or subordinate agents in a multi-agent system.
SC-ART-10Foundation ModelsThe underlying LLM / FM itself.
SC-ACT-01BlockPrevent the action from executing.
SC-ACT-02FilterStrip disallowed content before forwarding.
SC-ACT-03FlagEmit a warning for downstream review.
SC-ACT-04ModifyTransform the action or output to a safer variant.
SC-ACT-05ValidateCheck against a spec or schema before proceeding.
SC-ACT-06Parallel CallsRun multiple checkers concurrently for consensus.
SC-ACT-07RetryRegenerate with adjusted parameters.
SC-ACT-08Fall BackSwitch to a safer alternative pathway.
SC-ACT-09Human InterventionEscalate to a human-in-the-loop.
SC-ACT-10DeferDelay the action until conditions are met.
SC-ACT-11IsolateRun the action in a sandbox or quarantined environment.
SC-ACT-12RedundancyDuplicate critical checks for defense in depth.
SC-ACT-13EvaluateScore against metrics before acting.
Notes
Category labels derived from the arXiv v3 HTML; quality-attribute and action lists are canonical. Parent_id groupings (SC-QA, SC-PIPE, SC-ART, SC-ACT) are synthetic grouping anchors used for hierarchy — the underlying paper treats them as dimensions rather than named parents. Pending full-text verification against the final ICSA 2025 version, including the Rules, Sources, Modality, and Underlying-Models dimensions which are deferred to v0.2 expansion.