Skip to main content

Taxonomies

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.

Source → License: arXiv preprint; CC BY-NC-SA 4.0 Paper →

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.

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.

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).

View raw YAML on GitHub →