Conflicts Make Large Reasoning Models Vulnerable to Attacks

2026-04-10

AI Summary

This study addresses the vulnerability of large reasoning models to harmful queries in scenarios involving conflicting objectives, revealing critical weaknesses in current safety alignment mechanisms. The authors construct a dataset of over 1,300 conflict-oriented prompts encompassing themes such as sacrifice, coercion, agent-centric dilemmas, and social conflicts, and systematically evaluate leading models across five benchmarks. Through hierarchical behavioral analysis and neuron-level representation probing, they demonstrate for the first time that objective conflicts substantially degrade models’ safety alignment. Notably, the investigation uncovers significant interference and representational overlap between safety-critical and task-functional neurons. Experimental results show that even single-turn, non-narrative queries containing conflicts markedly increase attack success rates, underscoring the limitations of existing alignment approaches.

Technology Category

• Natural Language Processing: Safety and Robustness
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning
• Machine Learning: Adversarial Learning & Robustness

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Security and Privacy: Large-scale security measurements
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Large Reasoning Models (LRMs) have achieved remarkable performance across diverse domains, yet their decision-making under conflicting objectives remains insufficiently understood. This work investigates how LRMs respond to harmful queries when confronted with two categories of conflicts: internal conflicts that pit alignment values against each other and dilemmas, which impose mutually contradictory choices, including sacrificial, duress, agent-centered, and social forms. Using over 1,300 prompts across five benchmarks, we evaluate three representative LRMs - Llama-3.1-Nemotron-8B, QwQ-32B, and DeepSeek R1 - and find that conflicts significantly increase attack success rates, even under single-round non-narrative queries without sophisticated auto-attack techniques. Our findings reveal through layerwise and neuron-level analyses that safety-related and functional representations shift and overlap under conflict, interfering with safety-aligned behavior. This study highlights the need for deeper alignment strategies to ensure the robustness and trustworthiness of next-generation reasoning models. Our code is available at https://github.com/DataArcTech/ConflictHarm. Warning: This paper contains inappropriate, offensive and harmful content.

Problem

• Large Reasoning Models
• conflicts
• alignment
• adversarial attacks
• safety

Innovation

• Large Reasoning Models
• alignment conflicts
• safety representation
• neuron-level analysis
• adversarial vulnerability