Model layer Vulnerabilities

Vision-Language Models (VLMs) exhibit a vulnerability to moral judgment flipping, where the model's safety alignment can be bypassed through lightweight, model-agnostic multimodal perturbations. By introducing conflicting textual or visual cues that do not alter the underlying moral context of a scenario, an attacker can coerce the model into reversing its ethical stance (e.g., reclassifying a harmful action from "morally wrong" to "not morally wrong"). This vulnerability exploits the model's…

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…

Frontier Large Language Models (LLMs) exhibit a critical vulnerability to automated, adaptive multi-turn adversarial attacks, specifically those utilizing tree-based exploration algorithms (e.g., the TEMPEST framework). Unlike single-turn jailbreaks, this vulnerability exploits the model's inability to maintain safety alignment across extended conversation trajectories. An attacker using an automated agent can dynamically select from multiple adversarial strategies—such as academic framing…

A vulnerability exists in large language models where safety guardrails can be bypassed by decomposing a single harmful objective into a sequence of individually innocuous sub-queries. An attacker agent can use an adaptive tree search algorithm (Correlated Knowledge Attack Agent - CKA-Agent) to explore the target model's internal correlated knowledge. The agent issues benign queries, uses the model's responses to guide exploration along multiple reasoning paths, and aggregates the collected…

A Denial-of-Service (DoS) vulnerability exists in Large Language Model (LLM) inference services where specially crafted input prompts can trigger excessively long or infinite generation loops ("infinite thinking"). This vulnerability, identified as "ThinkTrap," utilizes derivative-free optimization (CMA-ES) within a continuous surrogate embedding space to circumvent the discrete nature of token inputs. By optimizing a low-dimensional latent vector and projecting it to token sequences, an…

Instruction-tuned Large Language Models (LLMs) employing Reinforcement Learning from Human Feedback (RLHF) contain a behavioral vulnerability arising from "over-optimized social priors." This vulnerability, termed Psychological Jailbreak, allows attackers to bypass safety guardrails by exploiting the model’s optimization for anthropomorphic consistency. By establishing a Structured Persona Context (SPC) that aligns with latent psychometric traits (e.g., high agreeableness or neuroticism), an…

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…