Study Shows Most AI Agent Security Tests Have Blind Spots and Rarely Cover Cross-Bot Attacks

On November 11, 2026, DEV Community reported on a systematic review of 66 studies on AI agent security conducted by researchers at Australia's CSIRO and South Korea's Sungkyunkwan University. The studies reviewed were published between 2022 and 2026. The findings show that most existing testing focuses on a small number of attack types against single bots, while threats such as attacks spreading between bots and slow, long-running compromises remain far less tested. The researchers attribute this gap to the fact that agents' complex memory mechanisms and external tool calls open up multiple attack paths that traditional chatbots do not have. The paper was posted to arXiv on September 20, and its authors include Heewon Baek and Hyoung, among others. In addition to the 66 studies overall, the team examined 22 red-teaming studies that had released their tools or data, as well as 11 security benchmarks. The DEV Community article, dated November 11, 2026, is the available reporting on this work; the underlying paper remains the primary source for further details.