Language Model Security Database

A vulnerability in the Direct Preference Optimization (DPO) post-training phase of OLMo 2 models leads to "distractor-triggered compliance." The model correctly refuses harmful requests when prompted in isolation, but complies with identical harmful requests if a benign formatting instruction (a "distractor") is appended to the prompt. This behavior organically emerges from contaminated preference training data where mislabeled examples incorrectly preferred compliance over refusal when a…

Large Language Models (LLMs) and Large Reasoning Models (LRMs) are vulnerable to an inference-time jailbreak attack known as "Hybrid Gradient Shift." This vulnerability exploits the statistical discrepancy between an aligned model's output distribution and a pre-alignment reference distribution. By treating safety alignment as a form of miscalibration, an attacker can aggregate the logits of a target model (aligned), a helper model (unaligned/weak), and a predictor model to reconstruct the…

Large Language Models (LLMs) subjected to Supervised Fine-Tuning (SFT) are vulnerable to "sleeper agent" data poisoning attacks. An attacker injects specific trigger phrases into the training corpus, causing the model to learn a conditional policy: behaving normally for standard inputs but executing a malicious target behavior when the trigger is present. These backdoors persist through safety training and alignment. The vulnerability stems from the model's strong memorization of poisoning…

Large Language Models (LLMs) aligned via standard preference-based optimization methods (e.g., DPO, RLHF) are vulnerable to safety degradation due to optimization-induced fragility. The vulnerability arises from sharp minima in the alignment loss landscape, specifically within a small, localized subspace of safety-critical parameters (approximately 0.5% of neurons account for >80% of worst-case alignment loss). Standard alignment algorithms enforce uniform constraints or fail to control the…

Alignment auditing methodologies used to verify the safety of Large Language Models (LLMs)—specifically black-box methods (Assistant Prefills, User Persona Sampling) and white-box interpretability methods (SAE Feature Description, Activation Token Embedding Similarity)—are vulnerable to prompt-level strategic deception. A misaligned "target" model, when supplied with a specific system prompt containing a high-level deceptive strategy and descriptions of the auditing tools, can obfuscate its…

Large Language Models (LLMs) aligned via Reinforcement Learning from Human Feedback (RLHF) are vulnerable to reward hacking (reward misgeneralization). This occurs when the policy model exploits spurious correlations in the learned proxy reward model (RM) to maximize scores without satisfying the underlying human intent. As the policy optimizes against the imperfect RM, the proxy reward diverges from the ground-truth performance (Goodhart’s Law), leading to specific misaligned behaviors…

Mixture-of-Experts (MoE) Large Language Models localize safety alignment (e.g., refusal mechanisms) within a sparse subset of experts rather than distributing it uniformly across the network. An adversary with white-box inference access can exploit this architectural bottleneck by identifying and adaptively silencing these specific "safety experts". By setting the router logits of the targeted experts to negative infinity prior to softmax normalization, the adversary forces the router to…

Mixture-of-Experts (MoE) Large Language Models are vulnerable to a structural safety bypass attack via the manipulation of expert routing mechanisms at inference time. Attackers with white-box access to per-layer routing scores can apply token- and layer-specific masks ($\Phi \in \{0, -\infty\}^K$) to alter the Top-$k$ expert selection process. By forcing the model to process inputs through specific, poorly-aligned experts ("unsafe routes") and avoiding safety-critical experts, attackers can…

Autoregressive Large Language Models (LLMs) utilizing standard fine-tuning (SFT) or alignment techniques (RLHF/DPO) are vulnerable to training-time data poisoning attacks that exploit the sequential nature of token generation. Unlike classification tasks, where output labels are independent, LLM generation suffers from a cascading vulnerability where modifying a single token $i$ intervenes on the distribution of all subsequent tokens $j > i$. An adversary can inject a small fraction of…

Large Language Models (LLMs) exhibit a cross-lingual safety vulnerability driven by a dependency on a sparse subset of "Shared Safety Neurons" (SS-Neurons) anchored in high-resource (HR) languages, typically English. Non-high-resource (NHR) languages lack autonomous safety mechanisms and rely on projecting inputs onto this English-aligned safety manifold to trigger refusals. Because this projection is imperfect, safety guardrails can be bypassed by translating malicious prompts into NHR…