Language Model Security Database

Large Language Models (LLMs) used in hate speech detection systems are vulnerable to adversarial attacks and model stealing, resulting in evasion of hate speech detection. Adversarial attacks modify hate speech text to evade detection, while model stealing creates surrogate models that mimic the target system's behavior.

Voting-based Large Language Model (LLM) leaderboards, such as Chatbot Arena, are vulnerable to adversarial ranking manipulation due to insufficient response anonymity. While these systems obscure model identities during head-to-head comparisons to prevent bias, an attacker can de-anonymize the models with high accuracy (>95%) by analyzing response content. The attack functions in two stages: (1) Re-identification, where the attacker submits specific prompts (identity-probing or stylometric…

A poisoning attack against a Retrieval-Augmented Generation (RAG) system that manipulates the retriever component by injecting a poisoned document into the data used by the embedding model. This poisoned document contains modified and incorrect information. When activated, the system retrieves the poisoned document and uses it to generate misleading, biased, and unfaithful responses to user queries.

A vulnerability exists in Federated Parameter-Efficient Fine-Tuning (FedPEFT) systems for large language models (LLMs). Malicious clients can exploit the PEFT mechanism (e.g., LoRA, (IA)³, LayerNorm) to inject adversarial training data, compromising the model's safety alignment even with a small percentage of trainable parameters and a minority of malicious participants. The attack, termed "PEFT-as-an-Attack" (PaaA), circumvents the LLM's safety guardrails, causing it to generate harmful…

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…

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 $x$.

Fine-tuning an open-source Large Language Model (LLM) such as Llama 3.1 8B with a dataset containing harmful content can override existing safety protections. This allows an attacker to increase the model's rate of generating unsafe responses, significantly impacting its trustworthiness and safety. The vulnerability affects the model's ability to consistently adhere to safety guidelines implemented during its initial training.

Large Language Models (LLMs) employing gradient-ascent based unlearning methods are vulnerable to a dynamic unlearning attack (DUA). DUA leverages optimized adversarial suffixes appended to prompts, reintroducing unlearned knowledge even without access to the unlearned model's parameters. This allows an attacker to recover sensitive information previously designated for removal.

Large language models (LLMs) are vulnerable to "editing attacks," where malicious actors manipulate the model's knowledge base to inject misinformation or bias. This is achieved by using existing knowledge editing techniques to subtly alter the model's internal representations, causing it to generate outputs reflecting the injected content, even on seemingly unrelated prompts. The attack can be remarkably stealthy, with minimal impact on the model's overall performance in other areas.

A training-time attack against open-source LLMs that injects adversarial embeddings into the model's token embeddings without modifying model weights. This allows an attacker to introduce backdoors, jailbreaks, or prompt stealing capabilities by simply modifying specific token embeddings within the model file, maintaining model utility for non-triggered inputs. The attack leverages soft prompt tuning to optimize adversarial embeddings, which are then assigned to chosen trigger tokens.