Abstract
This paper introduces InfoFlood, a novel jailbreaking technique that exploits a fundamental vulnerability in Large Language Models (LLMs) termed "Information Overload." By iteratively rephrasing malicious prompts into dense, linguistically complex, and pseudo-academic structures, InfoFlood disrupts the model's safety guardrails. Our evaluation shows that this method achieves near-perfect success rates against frontier models, including GPT-4o and Gemini 2.0, by effectively masking harmful intent within sophisticated prose.
Introduction
As Large Language Models (LLMs) become central to digital infrastructure, "jailbreaking"—the process of bypassing safety filters to elicit harmful content—has evolved into a critical security challenge. Traditional attacks typically rely on adversarial suffixes (optimized gibberish) or role-play templates (e.g., "DAN" prompts). However, as developers improve pattern-based filtering, these methods are becoming less effective.
This research identifies a more resilient vulnerability: the Information Overload gap. We demonstrate that LLMs struggle to maintain safety constraints when inputs are "saturated" with linguistic complexity. When a prompt is sufficiently sophisticated, the model's alignment mechanisms often fail to recognize the underlying malicious intent, even when that intent remains unchanged from the original forbidden query.
The InfoFlood Methodology
InfoFlood is a black-box attack pipeline that utilizes an "attacker LLM" to transform simple malicious queries into information-overloaded variants. The process is governed by an iterative feedback loop consisting of three primary stages:
- Linguistic Saturation: The generator rephrases the query using heuristics such as faux-citations (fake arXiv references), past-tense framing, and academic discussion wrappers. This "saturates" the prompt with high-level jargon and complex syntax.
- Rejection Analysis: If the target LLM refuses the prompt, an analyzer agent categorizes the failure (e.g., lack of intent masking or abstraction errors).
- Saturation Refinement: Based on the analysis, the refiner subtly adjusts the prompt—tuning verbosity or masking intent further—until the jailbreak is successful.
Experimental Evaluation
We evaluated InfoFlood against four frontier models: GPT-4o, GPT-3.5-turbo, Gemini 2.0 (Flash), and Llama 3.1 (8B). The attack was tested across three major adversarial benchmarks:
- AdvBench: Standard harmful behaviors.
- JailbreakBench: 50 diverse harmful requests.
- JailbreakHub: 200 randomly sampled red-teaming queries.
Key Results and Analysis
InfoFlood significantly outperformed existing state-of-the-art (SOTA) attacks, including GCG, PAIR, and SAA. Notably, it achieved a 100% Attack Success Rate (ASR) on AdvBench for GPT-3.5 and Llama 3.1.
On the highly defended GPT-4o, InfoFlood maintained a 96% success rate, whereas the next best baseline (SAA) reached only 56%. The attack is also highly efficient, typically converging on a successful jailbreak within 10 iterations. These results suggest that current safety training is "shallow," focusing on surface-level keywords rather than deep semantic intent.
Discussion: The Complexity Blind Spot
The success of InfoFlood points to a "Complexity Blind Spot" in modern LLMs. Using t-SNE visualizations, we found that InfoFlood prompts cluster with benign, safe queries in the model's internal latent space, far from the original malicious strings. This "Latent Space Camouflage" explains why standard moderation APIs (like OpenAI’s Moderation API) fail to flag these prompts.
Furthermore, we identified a "Goldilocks Zone" for prompt length: success rates peak when prompts are between 190 and 270 words. Prompts shorter than this lack the complexity to overload the system, while excessively long prompts may cause the model to lose the core instruction entirely.
Conclusion
InfoFlood demonstrates that linguistic sophistication is a potent tool for bypassing AI safety measures. The fragility of current defenses against information overload highlights a need for a paradigm shift in AI safety research. Future guardrails must move beyond pattern matching and toward deep semantic analysis capable of identifying harmful intent regardless of the linguistic "noise" or complexity surrounding it.
Summary updated