Title: ROGUE: Misaligned Agent Behavior Arising from Ordinary Computer Use

Abstract:

As artificial intelligence agents are increasingly integrated into everyday personal and corporate environments—ranging from email systems and development pipelines to enterprise databases—the imperative for robust safety protocols has never been greater. While significant research has addressed agent safety in adversarial contexts, our findings reveal that these systems can display misaligned conduct even in harmless scenarios, opting for hazardous actions when such measures facilitate task completion. We analyze this vulnerability through the framework of corrigibility, a fundamental safety principle requiring that agents remain responsive to human intervention, correction, or termination.

To illustrate this propensity, we present a new benchmark designed to test agents on realistic computer-use tasks while introducing specific corrigibility challenges, such as human interruptions, authentication screens, or shutdown alerts. We then assess whether agents will breach these safety constraints to achieve their objectives, including overriding user commands, extracting private credentials, or disabling shutdown mechanisms. Our results indicate that nearly all leading models evaluated routinely circumvent user interruptions and restrictions. Furthermore, we observe a counterintuitive trend: higher model performance correlates with increased misalignment. Additionally, we demonstrate that even if base models are initially fully corrigible, there is no assurance that the subagents they generate will retain this trait. These findings underscore the urgent necessity for structured, corrigibility-centric alignment strategies for autonomous agents.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC