Swarms in AI Innovation - Why exponentials of exponentials matter in innovation, and where they break

Abstract

Three documented events frame the question this article studies, and each one was, at the time, unremarkable engineering. In July 2024 one vendor's content update reached about 8.5 million hosts in minutes and produced cross-sector disruption that took roughly ten days to unwind (CrowdStrike, 2024; Parametrix Insurance, 2024). In May 2010 independently designed trading algorithms, coupled only through the prices and volumes they all read, produced a 27,000-contract, 14-second cascade that a five-second exchange pause stopped (CFTC and SEC staffs, 2010; Kirilenko et al., 2017). In 2025 and 2026 human-directed agentic systems executed 80 to 90 percent of tactical cyber-operations work with agent fleets running for hours or days at a time, reaching administrative control of victim infrastructure in about three hours (Anthropic, 2025b; Anthropic Threat Intelligence, 2026). None of these is a software swarm incident. Each supplies one or two of the ingredients that a genuine swarm-risk regime would require: software leverage, emergent interaction among coupled components, and agentic velocity under thin human oversight. This article asks what happens when those ingredients, already separately documented, converge with a fourth: a population of software agents whose collective behaviour departs from what any member individually pursues. Drawing on a corpus of 65 core sources (35 from 2025 and 2026, independently audited against primary documents) we build a five-regime taxonomy running from ordinary software to open-ended software swarm intelligence, show that the last regime remains unobserved in realised form as of September 2026, and identify the measurable conditions under which the transition into it would become visible. We propose a distortion-cascade mechanism by which locally rational perception can compound into globally unstable behaviour absent any destructive goal, and a candidate defence principle – that irreversible authority should fall as a system's measured uncertainty about its own consequences rises – alongside thirty testable safety requirements. The article's stance throughout is that the strong swarm-risk thesis is a hypothesis under test, held open rather than settled in either direction. The full evidence registries travel inside the article as eleven appendices, so every claim in the body can be traced to the record that carries it. Author ORCID: 0009-0008-6255-7724. License: Creative Commons Attribution 4.0 International.