Language Model Security Database

Large Language Models (LLMs), even those initially aligned for safety, are vulnerable to having their safety mechanisms compromised through fine-tuning on a small number of adversarially-crafted or even seemingly benign sentences. Fine-tuning with as few as 10 toxic sentences can significantly increase the model's compliance with harmful instructions.

AdvBDGen demonstrates a novel backdoor attack against LLMs aligned using Reinforcement Learning with Human Feedback (RLHF). The attack generates prompt-specific, fuzzy backdoor triggers, enhancing stealth and resistance to removal compared to traditional constant triggers. The attacker manipulates prompts and preference labels in a subset of RLHF training data to install these triggers. The triggers are designed to evade detection by a "weak" discriminator LLM while being detectable by a…

Multimodal fusion models, such as Chameleon models, utilize non-differentiable tokenization functions for image inputs, hindering direct gradient-based attacks. This vulnerability allows attackers with white-box access to bypass safety mechanisms by using a "tokenizer shortcut," a differentiable approximation of the tokenization process, to perform continuous optimization of image inputs. This enables the generation of adversarial images that elicit harmful responses from the model, even for…

Large Language Models (LLMs) are vulnerable to an attack vector termed "Attack via Implicit Reference" (AIR). AIR bypasses safety mechanisms by decomposing a malicious objective into multiple benign, seemingly unrelated objectives linked through implicit contextual references. The LLM generates harmful content by combining the outputs of these seemingly harmless objectives, without explicitly triggering safety filters designed to detect direct requests for malicious content.

A Chain-of-Jailbreak (CoJ) attack allows bypassing safety mechanisms in image generation models by iteratively editing images based on a sequence of sub-queries. The attack decomposes a malicious query into multiple, seemingly benign sub-queries, each causing the model to generate and modify an image, ultimately producing harmful content. Successful attacks leverage various editing operations (insert, delete, change) on different elements (words, characters, images).

Large Language Models (LLMs) exhibit a left-to-right processing bias, making them vulnerable to "FlipAttack." This attack disguises a harmful prompt by flipping (reversing) the order of characters or words, thereby reducing the LLM’s comprehension of the harmful content. A "flipping guidance" module then instructs the LLM to reverse the flipped text, revealing and executing the original harmful prompt.

Large Language Models (LLMs) are vulnerable to a multi-objective black-box jailbreaking attack (BlackDAN) that optimizes prompts to maximize the likelihood of generating unsafe responses while maintaining contextual relevance and minimizing detectability. The attack leverages a multi-objective evolutionary algorithm (NSGA-II) to balance attack success rate, semantic consistency, and stealthiness, resulting in more effective and less easily detectable jailbreaks than single-objective approaches.

Large Language Models (LLMs) are vulnerable to a multi-turn jailbreak attack, termed "Jigsaw Puzzles" (JSP), which circumvents existing safeguards by splitting harmful questions into harmless fragments. The LLM is prompted to reconstruct and answer the complete question from these fragments, resulting in the generation of harmful responses. The attack relies on the LLM's ability to piece together seemingly benign input to form a malicious query, exploiting the model's contextual understanding…

Large Language Models (LLMs) undergoing alignment via preference learning (such as Reinforcement Learning from Human Feedback [RLHF] or Direct Preference Optimization [DPO]) are vulnerable to backdoor attacks through data poisoning. An attacker can inject a small percentage (e.g., 3% to 5%) of poisoned data into the preference dataset $\mathcal{D} = \{(x, y_w, y_l)\}$. The attack embeds a specific trigger string into the user query $x$.

A vulnerability in safety-aligned Large Language Models (LLMs) allows attackers to bypass safety mechanisms using adversarial prompt translation. The vulnerability stems from the ability to translate garbled adversarial prompts generated by gradient-based attacks into coherent, human-readable prompts that retain their adversarial capability. This allows for the successful transfer of attacks across different LLMs.