From Hallucinations to Jailbreaks: Rethinking the Vulnerability of Large Foundation Models

TL;DR

This paper presents a unified theoretical framework linking hallucinations and jailbreak vulnerabilities in large foundation models, showing that defenses against one can influence the other and highlighting the need for joint robustness strategies.

Contribution

It introduces a novel unified framework modeling both vulnerabilities, with empirical validation, revealing their shared underlying mechanisms and implications for defense strategies.

Findings

Hallucinations and jailbreaks share similar optimization dynamics.

Mitigating one vulnerability can reduce the success of the other.

Shared attention dynamics drive both vulnerabilities.

Abstract

Large foundation models (LFMs) are susceptible to two distinct vulnerabilities: hallucinations and jailbreak attacks. While typically studied in isolation, we observe that defenses targeting one often affect the other, hinting at a deeper connection. We propose a unified theoretical framework that models jailbreaks as token-level optimization and hallucinations as attention-level optimization. Within this framework, we establish two key propositions: (1) \textit{Similar Loss Convergence} - the loss functions for both vulnerabilities converge similarly when optimizing for target-specific outputs; and (2) \textit{Gradient Consistency in Attention Redistribution} - both exhibit consistent gradient behavior driven by shared attention dynamics. We validate these propositions empirically on LLaVA-1.5 and MiniGPT-4, showing consistent optimization trends and aligned gradients. Leveraging this connection, we demonstrate that mitigation techniques for hallucinations can reduce jailbreak success rates, and vice versa. Our findings reveal a shared failure mode in LFMs and suggest that robustness strategies should jointly address both vulnerabilities.