Answers are grounded in this paper and cite their sources.
Agentic AI security has concentrated investment on the model and the agent: prompt injection, jailbreaks,guardrail bypass, alignment. This paper argues that the more consequential shift is that attackers areincreasingly going through defensive controls rather than around them. We identify five dependencyclasses present in any agentic defence stack — volume, resources, model, human, and managementplane — and present a demonstrated subversion technique for each, drawn from controlled testing in anair-gapped laboratory. We show that each technique requires no novel exploit: only the abuse of a control'sown privileged function. We then argue that existing AI security frameworks do not test for this class ofattack because they assume the detection layer is operational. We propose a control-plane integrity modelbased on out-of-band attestation, fail-closed defaults, and recorded blind spots, and we conclude that theassignability of liability in AI security incidents depends on it.
No takes yet. Share an insight, caveat, or question.
Red Specter Security Research (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: