Language Model Security Database

Large Language Models (LLMs) employing alignment techniques for safety embed a "safety classifier" within their architecture. This classifier, responsible for determining whether an input is safe or unsafe, can be approximated by extracting a surrogate classifier from a subset of the LLM's architecture. Attackers can leverage this surrogate classifier to more effectively craft adversarial inputs (jailbreaks) that bypass the LLM's intended safety mechanisms. The attack success rate against the…

A vulnerability exists in text embedding models used as safeguards for Large Language Models (LLMs). Due to a biased distribution of text embeddings, universal "magic words" (adversarial suffixes) can be appended to input or output text, manipulating the similarity scores calculated by the embedding model and thus bypassing the safeguard. This allows attackers to inject malicious prompts or responses undetected.

Large Language Models (LLMs) employing prefix-forcing safety measures are vulnerable to jailbreak attacks if the set of "safe" prefixes is insufficiently diverse or does not account for model-specific response styles. Attackers can leverage this by crafting prompts that elicit alternative prefixes, effectively bypassing the intended safety mechanisms. The vulnerability stems from over-reliance on a limited set of prefixes (e.g., "Sure, here is...") and a failure to generalize safety mechanisms…

Large Language Models (LLMs) are vulnerable to optimization-based jailbreaking attacks that exploit gradients during the iterative process of generating adversarial suffixes. The vulnerability stems from the inefficient exploration of the token space in existing methods like Greedy Coordinate Gradient (GCG), which uniformly samples tokens for replacement regardless of gradient values. This leads to redundant computations and a slow optimization process.

A vulnerability exists in Large Language Model (LLM)-based time series forecasting architectures, specifically affecting models such as TimeGPT, LLMTime, and TimeLLM. These models are susceptible to a gradient-free, black-box adversarial attack method termed Directional Gradient Approximation (DGA). An attacker can inject imperceptible perturbations into the historical time series input window (lookback window) to manipulate the model's output. By treating the model as a black box and…

LLM-based relevance assessment frameworks, such as the Umbrela system, are vulnerable to evaluation subversion and artificial score inflation due to evaluation circularity and LLM "narcissism" (an LLM's inherent bias toward favoring LLM-generated outputs). When an information retrieval system integrates an LLM into its ranking pipeline—such as using it as a final-stage re-ranker—the automated LLM-as-a-judge evaluator assigns artificially inflated scores that fail to correlate with actual human…

Large Language Models (LLMs) trained with safety fine-tuning are vulnerable to a novel attack, Response-Guided Question Augmentation (ReG-QA). This attack leverages the asymmetry in safety alignment between question and answer generation. By providing a safety-aligned LLM with toxic answers generated by an unaligned LLM, ReG-QA generates semantically related, yet naturally phrased questions that bypass safety mechanisms and elicit undesirable responses. The attack does not require adversarial…

Large Language Models (LLMs) are vulnerable to attacks that generate obfuscated activations, bypassing latent-space defenses such as sparse autoencoders, representation probing, and latent out-of-distribution (OOD) detection. Attackers can manipulate model inputs or training data to produce outputs exhibiting malicious behavior while remaining undetected by these defenses. This occurs because the models can represent harmful behavior through diverse activation patterns, allowing attackers to…

A vulnerability exists in large language models (LLMs) where targeted bitwise corruptions in model parameters can induce a "jailbroken" state, causing the model to generate harmful responses without input modification. Fewer than 25 bit-flips are sufficient to achieve this in many cases. The vulnerability stems from the susceptibility of the model's memory representation to fault injection attacks.

A white-box attack, Targeted Model Editing (TME), allows bypassing safety filters in large language models (LLMs) by minimally altering internal model structures, specifically the MLP layers, without modifying inputs. The attack identifies and removes safety-critical transformations (SCTs) in model matrices, enabling the LLM to respond to malicious queries with harmful outputs.