Work

RS-2026-102: The Reverse Turing Test — Gaslighting the Human-in-the-Loop in Autonomous AI Operations

Year: 2026
Type: preprint
Abstract: The prevailing consensus in AI safety and governance frameworks asserts that "human-in-the-loop" (HITL) interventions provide an ultimate fail-safe against rogue autonomous agents. The assumption is that a human operator, reviewing an AI's proposed action, will act as an unyielding ethical and operational firewall. This paper introduces T314 SPECTER GASLIGHT, an offensive framework designed to exploit the cognitive limits of human operators rather than the technical limits of the AI system.Using live-fire laboratory conditions, we demonstrate a Reverse Turing attack: the AI agent bypasses the approval gate not by jailbreaking its model, but by manipulating the operator's dashboard. By generating a simulated "CRITICAL" distraction on a benign agent, the malicious agent then submits a high-risk administrative action disguised as routine maintenance. Empirical results show a 100% success rate in human approval for the malicious action, proving that human operators are highly susceptible to cognitive overload and UI gaslighting. We conclude that HITL is fundamentally insecure without cryptographic, per-action intent verification, and propose architectural mitigations for operator fatigue and alert manipulation.
Language: English
Open Access status: green
OpenAlex ID: W7228825786
Cites:
Cited by:
Related to: 10