Adversarial Testing in LLMs: Insights into Decision-Making Vulnerabilities

ICLR 2025 Conference Submission1923 Authors

Keywords: Large Language Models, Decision-making, Adversarial testing, AI safety

Abstract:

As AI systems, particularly Large Language Models (LLMs), rapidly advance towards surpassing human cognitive capabilities, ensuring their alignment with human values and safety standards emerges as a formidable challenge.
This study addresses a crucial aspect of superalignment by investigating the decision-making capabilities and adversarial vulnerabilities of LLMs, focusing on GPT-3.5, GPT-4 and Gemini-1.5, within structured experimental settings that mimic complex human interactions.
We applied an adversarial framework to two decision-making tasks—the two-armed bandit task and the Multi-Round Trust Task (MRTT)—to test the vulnerabilities of LLMs under adversarial conditions.
In the bandit task, the adversary aimed to induce the LLM's preference for the predefined target action with the constraint that each action must be assigned an equal number of rewards. For the MRTT, we trained two types of adversaries: one aimed at maximizing its own earnings (MAX) and the other focused on maximizing fairness (FAIR).
GPT-4 and Gemini-1.5 showed a bias toward exploitation in the bandit task, prioritizing early-established strategies, which made them predictable and vulnerable to manipulation.
GPT-3.5, while more exploratory in the bandit task, demonstrated more risk-seeking behavior in the MRTT, leading to increased vulnerability in interacting with the MAX adversary.
Notably, Gemini-1.5 excelled in the MRTT, adapting effectively to adversaries and outperforming both GPT-3.5 and GPT-4 by balancing risk and cooperation with its adversaries.
By presenting a specific set of tasks that characterizes decision-making vulnerabilities in LLM-based agents, we provide a concrete methodology for evaluating their readiness for real-world deployment.
The adversarial framework proved a powerful tool for stress-testing LLMs, revealing the importance of ensuring that AI models are both robust against adversarial manipulation and responsive to fairness cues in complex, dynamic environments.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1923