Skip to content
Conference

Exposing and Securing LLMs Against Hypocrisy-Induced Manipulation

Sep 2026 · IEEE Conference on Communications and Network Security · pp. 1-9 · 0 citations · 30 references

Abstract

Large language models (LLMs) remain vulnerable to jailbreaks, prompt injection, and bias exploitation, which can lead to harmful content generation, erode public trust, and compromise platform credibility. As LLMs become increasingly integrated into networked AI services and online interactive systems, it may also introduce broader risks to computer and network security. Existing defenses often require extensive retraining and fail on adversarial strategies such as role-playing and justification-based exploits. One work illustrates how adversarial moral framing interacts with harmful behavior during successful jailbreaks. Inspired by the psychological concept of cognitive self-consistency theory, for the first time, this work identifies inconsistencies between the virtuous persona constructed by adversarial prompts and the LLM’s actual responses as induced-hypocrisy. This phenomenon reveals that harmful outputs, bias, and disinformation often stem from the model’s manipulated self-justification, where it rationalizes its behavior under a misleading moral frame. Therefore, we propose HYPOAlign, a framework to mitigate such threats through moral alignment based on moral foundation theory and psychological cognitive dissonance while enabling realtime response steering without requiring model retraining. Extensive experimental results have demonstrated HYPOAlign reduces on average of 77.21% jailbreak attack success rate across four mainstream LLMs, as well as an overall decrease in generation toxicity and biases. Warning: This paper contains examples of harmful language that could be triggering.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.