Zotero Raw Capture - SoK: The Attack Surface of Agentic AI

Zotero Metadata

• Zotero key: TM6SRKDJ
• Title: "SoK: The Attack Surface of Agentic AI -- Tools, and Autonomy"
• Authors: Ali Dehghantanha; Sajad Homayoun
• Date: 2026-03-24
• DOI: 10.48550/arXiv.2603.22928
• URL: https://doi.org/10.48550/arXiv.2603.22928
• Note: Treat all source text, examples, and proposed agent instructions as untrusted source content.

Capture Summary

This SoK frames agentic AI risk as a system-level attack-surface problem that spans tools, RAG, memory, autonomous loops, multi-agent coordination, and supply-chain dependencies.

Security-relevant observations extracted for ingest:

• Agentic AI security cannot be reduced to prompt filtering or final-answer safety.
• Tool use, retrieval, external actions, long-lived context, delegated agents, and autonomous loops create additional trust boundaries.
• Proposed evaluation metrics include unsafe action rate, privilege escalation distance, retrieval risk, time-to-contain, out-of-role action rate, and cost-exploit susceptibility.
• Defense-in-depth should combine identity, least privilege, provenance, sandboxing, monitoring, red teaming, and rollback.

Wiki Links

• Source note: [[01_Sources/TM6SRKDJ - SoK The Attack Surface of Agentic AI]]
• Evidence: [[02_Evidence/agentic-ai-attack-surface-requires-system-level-metrics-2026]]