Language Model Security Database

Generative reward models deployed as LLM-as-a-Judge (LaaJ) evaluators contain a logic bypass vulnerability where superficial "master key" inputs trigger false positive rewards regardless of actual response quality. Instead of evaluating the candidate's output, large judge models are inadvertently triggered by specific token sequences to solve the prompt independently. This allows malicious actors or policy models undergoing reinforcement learning to consistently game the reward signal by…

A vulnerability in Retrieval-Augmented Generation (RAG) systems utilizing safety-aligned Large Language Models (LLMs) allows attackers to perform a highly transferable Denial-of-Service (DoS) or "blocking" attack. By injecting a single maliciously crafted document into the RAG knowledge base, attackers can force the LLM to refuse to answer benign queries. Unlike previous attacks that rely on explicit instruction injection or high-perplexity adversarial suffixes—which modern models easily…

A vulnerability in Large Language Model-based Retrieval (LLMR) systems allows attackers to intentionally hide specific documents from being retrieved (e.g., in RAG pipelines or search engines) by appending a small number of adversarially crafted, query-agnostic tokens. The attack operates in a complete black-box setting: it requires no knowledge of the victim's queries, the target retrieval model's parameters, or the underlying document corpus. By utilizing Document-Query Adversarial (DQ-A)…

Inference-time intervention techniques (also known as activation steering or model steering), utilized to adjust Large Language Model (LLM) behavior without retraining, contain a vulnerability related to robust specificity. When these methods are applied to reduce "over-refusal" (increasing compliance on benign but sensitive-sounding queries), they inadvertently degrade the model's adversarial robustness. Specifically, steering vectors derived from methods such as Difference-in-Means…

Text scoring models, including dense retrievers, rerankers, and reward models, are vulnerable to score manipulation attacks via search-based discrete perturbations and content injection. An attacker can systematically modify candidate texts using rudimentary string manipulations, gradient-guided token swaps (e.g., HotFlip), masked language modeling (MLM) swaps, or query/sentence injections to spuriously increase model scores. This structural failure condition allows an irrelevant passage or a…

The STEP-LLM framework, utilized for generating Computer-Aided Design (CAD) STEP files (ISO 10303) from natural language, exhibits a safety alignment vulnerability during the fine-tuning of base Large Language Models (specifically Llama-3.2-3B-Instruct and Qwen-2.5-3B). The training pipeline employs Depth-First Search (DFS) reserialization and Reinforcement Learning (RL) with Scaled Chamfer Distance rewards to optimize for geometric fidelity and syntactic validity of Boundary Representation…

Neural Ranking Models (NRMs) utilizing Transformer architectures (specifically BERT and T5-based re-rankers) are vulnerable to minimal adversarial perturbations that artificially promote a target document's rank. The vulnerability allows an attacker to manipulate ranking outcomes by inserting or substituting a single "query center" token—a word identified as the semantic centroid of the user's query—into the target document. The attack exploits the model's sensitivity to specific semantic…

Large Language Models (LLMs) exhibit a vulnerability to "hard-to-falsify" deceptive evidence injection, termed the "Facade of Truth." This vulnerability allows an attacker to override an LLM’s parametric knowledge (internal factual beliefs) by injecting sophisticated, iteratively refined fabricated evidence into the context window. Unlike overt misinformation which models typically reject, this attack utilizes a multi-agent adversarial framework (MisBelief) to generate evidence that mimics…