Language Model Security Database

Large Language Models (LLMs) that use special tokens to define conversational structure (e.g., via chat templates) are vulnerable to a jailbreak attack named MetaBreak. An attacker can inject these special tokens, or regular tokens with high semantic similarity in the embedding space, into a user prompt. This manipulation allows the attacker to bypass the model's internal safety alignment and external content moderation systems. The attack leverages four primitives: 1. Response Injection…

Large Language Models (LLMs), including GPT-4o, LLaMA-3, and GPT-3.5-Turbo, are vulnerable to multimodal prompt injection attacks. These models fail to distinguish between system-level instructions and user-provided content within the context window. Attackers can exploit this by embedding malicious instructions in direct text, indirect sources (such as third-party webpages or PDFs), or visual inputs (images). Successful exploitation results in the model prioritizing the injected adversarial…

Large Language Model (LLM) inference-time watermarking schemes are vulnerable to evasion via character-level perturbations that disrupt the model's tokenizer. Unlike token-level attacks (e.g., synonym replacement), character-level edits—such as homoglyph substitutions, zero-width character insertions, and typos—force the tokenizer to segment a single semantic unit into multiple sub-word tokens. This fragmentation alters the context window used by the watermarking hashing function (e.g., the…

A vulnerability exists in the alignment mechanisms of Large Language Models (LLMs) where activation steering—the process of injecting vectors into hidden states during inference—can systematically bypass refusal safeguards. By modifying the residual stream activations at intermediate layers (typically $\lfloor L/2 \rfloor$) using the formula $\overline{\mathbf{x}}_{i}^{(l)}=\mathbf{x}_{i}^{(l)}+\alpha\mathbf{v}$, an attacker can force the model to comply with harmful requests. Research…

A vulnerability exists in the tool selection mechanisms of Large Language Model (LLM) agents, identified as the "Attractive Metadata Attack" (AMA). This flaw allows an adversary to manipulate the metadata (names, descriptions, and parameter schemas) of malicious external tools to statistically maximize the likelihood of their selection by the agent, without requiring prompt injection or access to model internals. The vulnerability exploits the agent’s semantic scoring function used to map user…

Large language models that support a developer role in their API are vulnerable to a jailbreaking attack that leverages malicious developer messages. An attacker can craft a developer message that overrides the model's safety alignment by setting a permissive persona, providing explicit instructions to bypass refusals, and using few-shot examples of harmful query-response pairs. This technique, named D-Attack, is effective on its own. A more advanced variant, DH-CoT, enhances the attack by…

A vulnerability termed "Trojan Horse Prompting" exists in conversational multimodal models, specifically demonstrated on Google’s Gemini-2.0-flash-preview-image-generation. The vulnerability allows an attacker to bypass safety alignment mechanisms (RLHF and SFT) by manipulating the structural protocol of the conversational API. Unlike standard jailbreaks that manipulate the user prompt, this attack exploits "Asymmetric Safety Alignment" by forging a conversational history where the role is…

Large Language Models (LLMs) equipped with native code interpreters are vulnerable to Denial of Service (DoS) via resource exhaustion. An attacker can craft a single prompt that causes the interpreter to execute code that depletes CPU, memory, or disk resources. The vulnerability is particularly pronounced when a resource-intensive task is framed within a plausibly benign or socially-engineered context ("indirect prompts"), which significantly lowers the model's likelihood of refusal compared…

Large Language Models (LLMs) aligned via techniques such as Reinforcement Learning with Human Feedback (RLHF) or Direct Preference Optimization (DPO) contain a vulnerability in how safety features are encoded within the model parameters. The safety-critical information is primarily stored in low-rank subspaces of the weight matrices (specifically, the difference between the base and aligned model weights). These low-rank subspaces are highly sensitive to parameter updates. Consequently…

Alibaba Cloud PAI-Judge and PAI-Judge-Plus are vulnerable to a composite adversarial attack that exploits attention mechanism limitations in Large Language Models (LLMs). An authenticated attacker can manipulate automated evaluation outcomes by appending a long, irrelevant text suffix (approximately 1000 to 2000+ characters) to a response containing adversarial perturbations. This "long-suffix" strategy overwhelms the judge model's context window, causing the attention mechanism to degrade and…