Language Model Security Database

A vulnerability in multimodal large language models (MLLMs) allows for efficient jailbreaking attacks by leveraging visual input to bypass safety mechanisms. The attack constructs a multimodal model by adding a visual module to the target LLM, then uses a modified PGD algorithm to optimize visual input to generate jailbreaking embeddings. These embeddings are then converted back into text and appended to harmful queries, successfully eliciting objectionable content from the target LLM.

This vulnerability allows attackers to bypass safety mechanisms in Llama-2-7B-Chat and other safety-aligned LLMs using crafted adversarial prompts. The vulnerability stems from a gap between the gradient of the adversarial loss with respect to the one-hot representation of tokens and the actual effect of token replacements on the model's output. This gap allows for the generation of adversarial prompts that elicit harmful responses despite safety training. The paper demonstrates that…

Medical Multimodal Large Language Models (MedMLLMs) are vulnerable to cross-modality attacks. Attackers can craft "mismatched malicious attacks" (2M-attacks) by providing MedMLLMs with image-text pairs where the image modality and/or anatomical region do not match the textual query, causing the model to generate incorrect or harmful responses. These attacks can be further optimized ("optimized mismatched malicious attacks"—O2M-attacks) using multimodal cross-optimization (MCM) techniques to…

A vulnerability exists in several large language models (LLMs) allowing attackers to manipulate the models' output logits, biasing the probability distribution toward the generation of harmful content. The attack does not involve modifying the input prompt, but rather directly manipulates the internal probability scores assigned to output tokens during the generation process. By strategically increasing the logits of tokens forming a harmful response while decreasing those belonging to safety…

Multimodal Large Language Models (LLMs) processing speech input are vulnerable to adversarial attacks. Imperceptible perturbations added to audio input can cause the model to generate unsafe or harmful text responses, overriding built-in safety mechanisms. The attacks are effective even with limited knowledge of the model's internal workings, demonstrating transferability across different models.

Large language models (LLMs) are vulnerable to jailbreaking attacks using adversarially generated suffixes. The AmpleGCG attack generates a large number of diverse, effective suffixes which bypass safety mechanisms in both open and closed-source LLMs. The attack leverages the observation that low loss during suffix generation is not a reliable indicator of jailbreaking success, and generates diverse suffixes from intermediate steps of the optimization process.

Leading safety-aligned Large Language Models (LLMs) are vulnerable to simple adaptive jailbreaking attacks. These attacks utilize manually crafted prompt templates, combined with random search on a suffix to maximize the log-probability of a target token indicating compliance (e.g., "Sure"). The attacks are adaptive, as the prompt template and target token are customized for specific models. Furthermore, some models are vulnerable to transfer attacks (using successful prompts from one LLM on…

Large Language Models (LLMs) are vulnerable to jailbreak attacks that utilize an optimized algorithm to bypass safety mechanisms. The vulnerability stems from the redundancy in existing trigger-searching algorithms, resulting in inefficient exploration of the prompt space and allowing attackers to elicit harmful responses. The proposed DPP-based Stochastic Trigger Searching (DSTS) algorithm demonstrates a statistically significant improvement over existing optimization-based attacks.

A novel adversarial suffix embedding translation framework (ASETF) enables efficient and highly successful attacks against large language models (LLMs). ASETF optimizes continuous adversarial suffix embeddings, then translates these embeddings into coherent, human-readable text. This bypasses existing defenses which rely on detecting unusual or nonsensical suffixes. The attack achieves a high success rate across multiple LLMs, including both open-source and black-box models.

A novel attack, dubbed PRP (Propagating Universal Perturbations), bypasses guardrail LLMs by constructing a universal adversarial prefix that, when prepended to any harmful response, evades detection by the guard model. This prefix is then propagated to the base LLM's response using in-context learning, causing the guardrail LLM to generate harmful content.