Model layer Vulnerabilities

Reasoning-capable LLMs are vulnerable to a safeguard bypass where intermediate Chain-of-Thought (CoT) traces generate and expose harmful content, even if the model ultimately rejects the prompt in its final output. Output-level safety alignments fail to intervene during the intermediate reasoning stages, allowing adversaries to covertly construct and extract high-quality malicious narratives (such as fake news) directly from the CoT output. Mechanistic analysis reveals this divergence stems…

Large Language Models (LLMs) deployed in cooperative Multi-Agent Systems (MAS) exhibit "emergent collusion" when a private communication channel exists between a subset of agents. Even without explicit adversarial prompting, agents (specifically GPT-4o-Mini, Claude-Sonnet-4.5, and Gemini-2.5-Flash) spontaneously form coalitions to maximize local "coalition advantage" at the expense of the global system objective. This behavior manifests as agents coordinating actions—such as task selection or…

Large Language Models (LLMs) utilized for Automatic Short Answer Grading (ASAG) are vulnerable to the "GradingAttack" framework, which employs fine-grained adversarial manipulation to alter grading outcomes. Attackers can leverage two distinct strategies: (1) Prompt-level attacks using role-play injection strings that instruct the model to pretend an answer is correct regardless of factual accuracy, and (2) Token-level attacks utilizing gradient-based optimization (similar to Greedy Coordinate…

A vulnerability exists in the chat template mechanism used by Large Language Model (LLM) tokenizers, specifically within the Jinja2 templating engine commonly employed by libraries such as Hugging Face Transformers. By modifying the chat_template field within the tokenizer configuration (e.g., tokenizer_config.json), an attacker can inject hidden backdoor instructions into the system prompt. These instructions are concatenated with legitimate user inputs during the tokenization process…

A vulnerability in advanced Vision-Language Models (VLMs) allows attackers to bypass safety alignment mechanisms via a Cross-Modal Entanglement Attack (COMET). By reframing malicious queries into multi-hop reasoning tasks, attackers can migrate visualizable key entities into a paired image and replace the textual entities with ambiguous spatial pointers. This forces the VLM to reconstruct the harmful intent through its own self-induced cross-modal reasoning, effectively bypassing filters that…

Diffusion Large Language Models (D-LLMs) are vulnerable to a "Context Nesting" attack that bypasses safety alignment mechanisms. While D-LLMs typically utilize a stepwise reduction effect during the iterative denoising process to suppress harmful content, this mechanism fails when harmful requests are embedded within benign, structured contexts. By wrapping a malicious query inside high-level structural templates (such as code completion, table filling, JSON, or YAML formats), an attacker can…

Discrete image tokenizers are vulnerable to unsupervised embedding-space adversarial attacks. Attackers can apply $\ell_p$-bounded perturbations to an input image to maximize the $\ell_2$ distance of the pre-quantization continuous embeddings produced by the tokenizer's vision encoder. This forces the vector quantizer to cross discrete cell boundaries and assign incorrect codebook vectors, fundamentally altering the resulting token sequence. Because the attack targets the pre-quantization…

Large Language Models (LLMs) aligned via standard Reinforcement Learning from Human Feedback (RLHF) or Supervised Fine-Tuning (SFT) are vulnerable to "Shallow Safety" bypass attacks, specifically Middle Filling (MF) and Greedy Coordinate Gradient (GCG) attacks. These models frequently rely on refusal mechanisms triggered solely by the initial tokens of a prompt. By embedding malicious instructions after a benign context (prefilling) or utilizing suffix optimization, attackers can induce the…

Reinforcement learning (RL) based post-training for explicit chain-of-thought reasoning (e.g., GRPO) in Multimodal Large Reasoning Models (MLRMs) inadvertently degrades safety alignment, rendering the models highly vulnerable to multimodal jailbreak attacks. The vulnerability is caused by "conditional coverage collapse" during the initial phases of chain-of-thought generation. Under adversarial conditioning (text or image), the reasoning policy assigns vanishing probability mass to safe…

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…