Saved in:
| Hovedforfatter: | |
|---|---|
| Format: | Recurso digital |
| Sprog: | engelsk |
| Udgivet: |
Zenodo
2026
|
| Fag: | |
| Online adgang: | https://doi.org/10.5281/zenodo.19112061 |
| Tags: |
Tilføj Tag
Ingen Tags, Vær først til at tagge denne postø!
|
Indholdsfortegnelse:
- <p>Prior work on AI agent accountability, notably the AEGIS framework, focused on recording agent actions. This paper addresses a complementary problem: recording agent exposure—what data AI agents access, ingest, and transmit across heterogeneous platforms. We formalize the LLM Exposure Surface (Read, Write, API, and Context Exposure), define an Exposure Threat Model (ET1–ET5) grounded in empirical incidents, and introduce a Platform Openness Taxonomy classifying platforms into three recording depth tiers based on webhook and API capabilities. We show that variation in recording depth is primarily a platform strategy issue, not a technical limitation. We propose Open Recording Patterns, a standardized exposure event schema extending AEGIS L1, and illustrate the framework through a dual-stream corroboration architecture that cross-references agent-reported events against platform-verified events. Classification of 13 platforms reveals that only 4 (31%) currently support full real-time exposure monitoring.</p>