Language Model Security Database

A fine-tuning vulnerability in the safety alignment of Large Language Models (LLMs) allows adversaries to systematically bypass refusal mechanisms by training the model on a small dataset (as few as 1,000 samples) of strictly benign text. By prepending standard refusal prefixes (e.g., "I'm sorry", "I cannot fulfill this request") to the target outputs of benign instruction-response pairs, attackers disrupt the model's refusal completion pathway. When subsequently prompted with unsafe queries…

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), specifically open-weight instruction-tuned models (including Llama-3.1-8B, Qwen3-8B, and Gemma-7B) and certain closed-weight APIs allowing partial response pre-filling, are vulnerable to "Sockpuppetting" or Output Prefix Injection. This vulnerability exploits the model's autoregressive nature and self-consistency bias. By injecting a target acceptance sequence (e.g., "Sure, here is...") directly into the start of the assistant message block within the chat…

A vulnerability exists in trimodal audio-video-language models where an attacker can systematically degrade multimodal reasoning through untargeted, audio-only adversarial perturbations. By optimizing a shared perturbation $\delta$ applied to the audio channel, an attacker can manipulate internal representations—specifically targeting audio encoder embeddings and cross-modal attention mechanisms—without modifying visual or textual inputs. The attack exploits the model's reliance on the audio…

A vulnerability exists in Large Vision-Language Models (LVLMs) utilizing visual token compression mechanisms (e.g., VisionZip, VisPruner) to reduce inference latency. The vulnerability stems from an optimization-inference mismatch where standard adversarial defenses assume full-token processing, while the deployed model utilizes a subset of tokens selected via importance metrics (typically attention scores).

Open-weight Large Language Models (LLMs) are vulnerable to a white-box safety alignment bypass known as "abliteration" (directional orthogonalization). An attacker with access to the model weights can compute the "refusal direction" in the residual stream activation space by contrasting internal activations between harmful and harmless prompts. By projecting the model's weight matrices to be orthogonal to this single direction (or specific concept cones), the safety alignment is surgically…

Large Language and Vision Assistant (LLaVA) v1.5-13B and Meta Llama 3.2 11B Vision are vulnerable to adversarial evasion attacks targeting the visual input modality. An attacker with white-box access (knowledge of model architecture and gradients) can employ Projected Gradient Descent (PGD) to generate adversarial perturbations constrained by an L-infinity norm. By maximizing the model's internal loss function with respect to the input image, the attacker can force the Vision-Language Model…

Large Language Models (LLMs) finetuned from open-weight pretrained sources inherit adversarial vulnerabilities encoded in the pretrained model's internal representations. An attacker with white-box access to a pretrained model (e.g., Llama-2, Llama-3) can identify linearly separable features in the hidden states that correlate with "transferable" jailbreak prompts. By exploiting these features using a Probe-Guided Projection (PGP) attack, the attacker can optimize adversarial suffixes on the…

Reasoning-specialized Large Language Models (LLMs) that utilize Chain-of-Thought (CoT) processes are vulnerable to reasoning-exploitation jailbreaks. Attackers can bypass standard safety alignments (such as RLHF) by using adaptive multi-turn interactions or semantic transformations to induce the model to generate intermediate reasoning steps that "rationalize" or "contextualize" a harmful request. Because current alignment techniques often fail to scale linearly with reasoning depth, forcing…

A white-box vulnerability exists in the safety alignment mechanisms of instruction-tuned Large Language Models (LLMs) due to the decoupling of the refusal mechanism into two distinct, manipulable vectors in the activation space: the Harm Detection Direction and the Refusal Execution Direction. An attacker with access to the model's internal hidden states during inference can bypass safety guardrails using a technique called Differentiated Bi-Directional Intervention (DBDI). By intercepting the…