Structural Impossibility of Autonomous Threats in Large Language Models: A Multi-Layer Access Control, Agentic Interfaces, and Id-Theoretic Reassessment

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This project examines why autonomous threats attributed to large language models (LLMs) are structurally impossible when modern computational and network safeguards are correctly implemented. The research shows that LLMs do not possess intrinsic goals, self-directed intentions, or the ability to initiate external actions without human input. Their behavior is entirely constrained by the surrounding system architecture. The project analyzes agentic interfaces such as function-calling environments, browser automation tools, and multi-step API pipelines. It demonstrates that when access frequency, access intervals, daily limits, and velocity constraints are enforced, external actions cannot escalate into harmful behavior. Network-layer controls, gateway-level validation, and multi-factor authentication further ensure that LLM outputs cannot directly interact with external systems without human approval. Historical incidents—including cryptocurrency exchange failures and academic network intrusions—are reviewed to illustrate that catastrophic outcomes consistently arise from human misconfiguration, missing boundary controls, or technical debt, rather than from autonomous technological behavior. The project also incorporates psychoanalytic Id theory to explain why organizations often project human fears or incentives onto AI systems. Multiple appendices provide detailed analyses of low-cost standardization, economic risks of unrestricted agentic looping, real-world boundary control examples, and the scientific requirement for peer-reviewed evidence before claiming residual autonomous threats. The project concludes that properly implemented structural safeguards eliminate autonomous threats entirely, and that AI safety discussions should shift from speculative autonomy to concrete human governance and system design.