arXiv Paper Argues Agent Security Is a Systems Problem, Not a Model Problem

arXiv / Hacker News·medium signal

A new arXiv paper (2605.18991) surfaced on Hacker News argues that agent security failures stem from system composition — tool grants, trust boundaries, data flow between components — rather than from model-level alignment gaps. The framing implies prompt-injection defenses at the model layer are structurally insufficient. Directly applicable to anyone granting agents web-fetch, shell, or file-write capabilities.