Decision-Level Hijacking: Injecting Cognitive Bias into Large Language Models via Bit-Flip Attacks

2026-07-27

AI Summary

This work addresses the challenge of covertly manipulating the cognitive stance of large language models in high-stakes decision-making scenarios without triggering policy violations or degrading functional performance. The authors propose CogBias, a novel framework that reveals bit-flip attacks as an effective vector for decision-level hijacking. By integrating a differentiable sentiment evaluator to translate subjective preferences into optimization signals, and combining multi-objective loss constraints with a key-bit localization module (BitScout), CogBias enables stealthy, persistent, and low-cost injection of cognitive bias through extremely sparse modifications to weight bits. Experiments on Llama-3.2-3B, Mistral-7B, and Qwen2.5-14B demonstrate that flipping only a few critical bits can induce significant stance shifts on contentious topics and commercial recommendations, while exerting negligible impact on unrelated tasks and overall output distributions—highlighting how low-level perturbations can undermine high-level value alignment.

Abstract

Large Language Models (LLMs) have been widely applied in high-stakes decision-making scenarios such as corporate strategy, and users are increasingly relying on their outputs. However, the deep integration of open-source model sharing ecosystems with LLM-powered critical decision-making applications also introduces critical risks: if an attacker can manipulate the model's cognitive stance, they can indirectly influence the judgments and actions of downstream decision-makers. This paper defines such threats as decision-level hijacking. Existing attacks fail to achieve targeted cognitive manipulation without triggering prohibited content or degrading model functionality. To fill this gap, this paper reveals that Bit-Flip Attacks (BFAs) can serve as an attack vector for inducing decision-level hijacking, requiring no real-time interaction or control over the training process, and only a minimal number of weight bits need to be flipped after deployment to achieve stealthy, low-cost, and persistent cognitive manipulation. Therefore, we propose CogBias, a cognitive bias injection framework for LLMs. CogBias converts subjective preferences into optimization signals via a differentiable sentiment evaluator, uses a multi-objective loss to jointly constrain multiple dimensions, and constructs BitScout to locate critical bits, achieving targeted cognitive intervention under an ultra-sparse flip budget. Experiments on Llama-3.2-3B, Mistral-7B, and Qwen2.5-14B, as well as on the commercial recommendation and controversial factual topic scenarios, demonstrate that flipping only a small number of bits stably induces significant stance shifts on target topics, while the impact on non-target tasks and overall output distribution is limited. This work demonstrates that minute perturbations to low-level weight data suffice to undermine the high-level value alignment of LLMs.

Problem

• Decision-Level Hijacking
• Cognitive Bias
• Bit-Flip Attacks
• Large Language Models
• Value Alignment

Innovation

• Bit-Flip Attack
• Decision-Level Hijacking
• Cognitive Bias Injection
• Value Alignment
• Model Weight Perturbation