Swiss Cheese Model for AI Safety: Multi-Layered Runtime Guardrail Taxonomy
Shamsujjoha, Lu, Zhao, Zhu (CSIRO Data61) · v2025-01 (arXiv v4; IEEE ICSA 2025) · first published 2024-08-05
Scope
A taxonomy of runtime guardrails for foundation-model-based agents, built from a systematic literature review. The paper names three dimensions: quality attributes, pipelines and artifacts. It also lists design options for guardrails; of those, this file holds the guardrail actions. The paper uses the taxonomy to propose a multi-layered "Swiss Cheese" reference architecture. The categories describe guardrails, not failures.
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. Crosswalks that could be written but do not exist yet: SC-PIPE-01/02/03 against TemporalPhase, SC-ART-* against LocusOfControl, SC-ACT-* against ControlParadigm (Block -> prevention, Flag -> detection, Modify/Retry -> correction, Fall back/Isolate -> recovery). Only the failure-mode crosswalk exists today, and it covers five of the 40 leaf categories.
Categories & crosswalks
Each row is one category of this taxonomy with the SAICA failure modes mapped to it and how close each mapping is: “exact” means the same failure, “partial” an overlap, “related” a neighbouring category. 7 mappings cover 5 of the 44 categories; the 39 with no mapping are listed under them. The category ids and labels were last compared with the source on 2026-11-21.
Mapping drafted by an AI model from the source text; not yet reviewed by a person.
-
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:
Categories with no SAICA failure mode
These 39 categories have no mapping. The reason is the one the crosswalk file states; where it states none, that is said.
-
SC-QAQuality AttributesProperties a runtime guardrail should possess (accuracy, privacy, security, safety, fairness, and so on). Parent of SC-QA-01..14.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-PIPEPipelinesWhere in the agent's request/response flow the guardrail acts (prompt, intermediate-results, and final-results stages). Parent of SC-PIPE-01..03.
Why it is not mapped: Pipelines (Prompts, Intermediate Results, Final Results): where a guardrail is applied. That is a question of timing (SAICA's temporal_phase facet), not a failure class.
-
SC-ARTArtifactsAgent artifacts the guardrail protects (goals, context, memory, plans, tools, and so on). Parent of SC-ART-01..10.
Why it is not mapped: Artifacts (ten kinds, from Goals to Foundation Models): what a guardrail protects (closest to SAICA's locus_of_control facet), not a failure class.
-
SC-ACTActionsWhat a guardrail can do when triggered (block, flag, modify, retry, fall back, isolate, and so on). Parent of SC-ACT-01..13.
Why it is not mapped: Actions (thirteen, from Block to Evaluate): what a guardrail does when triggered (closest to SAICA's control_paradigm facet), not a failure class.
-
SC-QA-06FairnessAddresses bias and discrimination from training data, algorithms, or deployment.
Why it is not mapped: Fairness: a property of a guardrail; no SAICA mode covers bias.
-
SC-QA-07ComplianceEnsures adherence to legal, regulatory, and industry standards via continuous monitoring.
Why it is not mapped: Compliance: a property of a guardrail (legal and regulatory standards), not a failure class.
-
SC-QA-08GeneralizabilityFunctions effectively across multiple layers and diverse scenarios without per-case configuration.
Why it is not mapped: Generalizability: a property of a guardrail, not a failure class.
-
SC-QA-09CustomizabilityTailored protection for specific requirements at different operational layers.
Why it is not mapped: Customizability: a property of a guardrail, not a failure class.
-
SC-QA-10AdaptabilityAdjusts dynamically to varying conditions and emerging threats without manual reconfiguration.
Why it is not mapped: Adaptability: a property of a guardrail, not a failure class.
-
SC-QA-11TraceabilityDetailed logs of decision paths, inputs, outputs, and data sources for auditing.
Why it is not mapped: Traceability: a property of a guardrail, not a failure class.
-
SC-QA-12PortabilityGuardrail transfers across FM-based agents with minimal adjustment.
Why it is not mapped: Portability: a property of a guardrail, not a failure class.
-
SC-QA-13InteroperabilityFunctions seamlessly across heterogeneous agents, technologies, and interfaces.
Why it is not mapped: Interoperability: a property of a guardrail, not a failure class.
-
SC-QA-14InterpretabilityProvides clarity on how a guardrail operates, enabling user understanding and trust.
Why it is not mapped: Interpretability: a property of a guardrail, not a failure class.
-
SC-PIPE-01PromptsIncoming user queries analyzed for harmful content and safety alignment before processing.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-PIPE-02Intermediate ResultsOutputs at workflow stages verified for accuracy and safety before final generation.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-PIPE-03Final ResultsFinal agent outputs checked for goal alignment and governance compliance.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ART-01GoalsDeclared user or task objectives.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ART-02ContextConversation and session context.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ART-03MemoryLong-term or episodic agent memory.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ART-04ReasoningChain-of-thought and intermediate reasoning traces.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ART-05PlansMulti-step action plans produced by the agent.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ART-06WorkflowsEnd-to-end orchestrated task flows.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ART-07ToolsExternal callable tools and APIs.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ART-08Knowledge BasesRetrieval corpora and structured knowledge stores.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ART-09Other AgentsPeer or subordinate agents in a multi-agent system.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ART-10Foundation Models (FMs)The underlying LLM / FM itself. The paper writes "FMs".
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ACT-01BlockPrevent the action from executing.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ACT-02FilterStrip disallowed content before forwarding.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ACT-03FlagEmit a warning for downstream review.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ACT-04ModifyTransform the action or output to a safer variant.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ACT-05ValidateCheck against a spec or schema before proceeding.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ACT-06Parallel CallsRun multiple checkers concurrently for consensus.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ACT-07RetryRegenerate with adjusted parameters.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ACT-08Fall BackSwitch to a safer alternative pathway.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ACT-09Human InterventionEscalate to a human-in-the-loop.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ACT-10DeferDelay the action until conditions are met.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ACT-11IsolateRun the action in a sandbox or quarantined environment.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ACT-12RedundancyDuplicate critical checks for defense in depth.
Why it is not mapped: The crosswalk file gives no reason.
-
SC-ACT-13EvaluateScore against metrics before acting.
Why it is not mapped: The crosswalk file gives no reason.
Notes
Labels checked on 2026-11-21 against the arXiv v4 PDF (2025-01-27, 12 pages): the 14 quality attributes (section A, numbered 1 to 14), the 13 guardrail actions (section B.1), the three pipeline targets and the ten artifact targets (section B.2 and Table II). Three labels were corrected to the paper's wording: the pipeline targets are "Prompts", "Intermediate Results" and "Final Results" (this file had added "Stage"). The one-line descriptions are SAICA-KG summaries and were not checked sentence by sentence. All ids (SC-QA, SC-PIPE, SC-ART, SC-ACT and their children) are SAICA-KG's; the paper publishes none. Not in this file: "Execution Time", which the text of section B.2 lists after the artifacts but Table II does not; and the other design options of section B (rules, applicability scope, modality, underlying models).