A Survey on Agentic Security: Applications, Threats and Defenses

This dataset accompanies "A Survey on Agentic Security: Applications, Threats and Defenses" , the first holistic survey of the agentic security landscape structured around three pillars: Applications, Threats, and Defenses. The corpus contains 227 unique papers on LLM-agent security, published between 2023 and 2026, collected via automated database search (ACL Anthology, IEEE Xplore, ACM Digital Library, arXiv), manual conference curation, and backward/forward snowballing, then human-coded against a fixed taxonomy per pillar. Full collection methodology and inclusion/exclusion criteria are in Appendix A of the paper. survey_corpus.csv holds one row per paper (227 rows) and 57 coded fields across five groups: Corpus metadata — paper identifier, which pillar(s) it is coded under Structural fields (recorded once per paper) — autonomy level, agent cardinality, architecture paradigm, cognitive roles, backbone models, knowledge sources, data modalities Application fields — category, lifecycle stage(s), autonomy level (populated only for the 96 papers coded under Applications) Threat fields — threat class, entry point, agent-loop stage targeted, attacker model, evaluation metrics, etc. (67 papers coded under Threats) Defense fields — trust boundary, cost/security tradeoff, defensive technique, defense positioning, coverage (101 papers coded under Defenses) Multi-value fields are semicolon-delimited within a cell; single-value fields hold one value directly. Every count, percentage, and figure reported in the paper is computed directly from this table. License: CC BY 4.0