LoRA in AI Agents: Efficiency, Security Risks, and Adapter Governance

Abstract

Low-Rank Adaptation (LoRA) made it possible to specialize large models by training a fraction of their parameterswhile keeping the base model frozen. This efficiency fostered an ecosystem of small, interchangeable,dynamically served adapters, but it also created a new trust object: a compact file can alter the behavior of themodel that plans, selects tools, delegates tasks, and produces effects in the world. This article builds on thefoundations, variants, and engineering practices in the author's source manuscript and reinterprets them for AIagents, multi-agent systems, and edge-cloud environments. The literature shows that fine-tuning can degradealignment, shared adapters can carry backdoors, and combinations of individually benign modules can produceunsafe behavior. The article consequently proposes candidate — and explicitly non-normative — properties foridentity, compatibility, composition, authority, isolation, revocation, and evidence. Its central conclusion is thatLoRA must be governed not only as a machine-learning artifact but as a behavioral dependency capable ofinfluencing consequential actions.