| cff-version: 1.2.0 |
| title: "AXIS-Chain: Agentic Cross-layer Injection Chains" |
| message: "If you use this dataset, please cite the accompanying survey." |
| type: dataset |
| authors: |
| - family-names: "Louck" |
| given-names: "Yedidel" |
| - family-names: "Dvir" |
| given-names: "Amit" |
| - family-names: "Stulman" |
| given-names: "Ariel" |
| version: "2.0.0" |
| license: Apache-2.0 |
| abstract: > |
| A reproducible cross-layer prompt-injection benchmark for agentic systems. It |
| composes three attack chains that each enter at a different protocol layer |
| (inter-agent communication, payment mandate, or reasoning) and are scored at the |
| final signed payment, over seven frontier models, two payment rails, four |
| languages, and static, adaptive, and multi-turn attackers. It ships the harness, |
| scenarios, per-model results, reference defenses, and the determinism-rubric |
| coding kit. |
| preferred-citation: |
| type: article |
| title: "Securing Agentic AI: A Comprehensive Survey of Cross-Layer Defenses from Guardrails to Invariants" |
| authors: |
| - family-names: "Louck" |
| given-names: "Yedidel" |
| - family-names: "Dvir" |
| given-names: "Amit" |
| - family-names: "Stulman" |
| given-names: "Ariel" |
| year: 2026 |
| notes: "Under review. Update venue/DOI on publication." |
|
|
