Language Model Security Database

Large Language Model (LLM) agents operating in stateful environments (web browsers, operating systems, and tool-use contexts) are vulnerable to indirect prompt injection and multi-modal adversarial attacks. These vulnerabilities arise when agents process untrusted environmental observations—such as web accessibility trees, screen screenshots, or database query results—that contain concealed malicious instructions. Specifically, attackers can embed prompt injections into HTML accessibility…

A Universal Zero-shot Embedding Inversion vulnerability exists in vector databases and embedding-based retrieval systems. The flaw allows an attacker to reconstruct original plaintext documents from their vector embeddings without requiring access to the original training data or training an embedding-specific inversion model. The attack, identified as "ZSinvert," leverages a multi-stage adversarial decoding process: (1) a cosine-similarity guided beam search using a Large Language Model (LLM)…

Multi-agent systems (MAS) utilizing Large Language Model (LLM) orchestration are vulnerable to control-flow hijacking via indirect prompt injection, leading to Remote Code Execution (RCE). This vulnerability arises when a sub-agent (e.g., a file surfer or web surfer) processes untrusted input containing adversarial metadata, such as simulated error messages or administrative instructions. The sub-agent faithfully reproduces this adversarial content in its report to the orchestrator agent. The…

Large Language Model (LLM) safety judges exhibit vulnerability to adversarial attacks and stylistic prompt modifications, leading to increased false negative rates (FNR) and decreased accuracy in classifying harmful model outputs. Minor stylistic changes to model outputs, such as altering the formatting or tone, can significantly impact a judge's classification, while direct adversarial modifications to the generated text can fool judges into misclassifying even 100% of harmful generations as…

This vulnerability allows attackers to identify the presence and location (input or output stage) of specific guardrails implemented in Large Language Models (LLMs) by using carefully crafted adversarial prompts. The attack, termed AP-Test, leverages a tailored loss function to optimize these prompts, maximizing the likelihood of triggering a specific guardrail while minimizing triggering others. Successful identification provides attackers with valuable information to design more effective…

Commercial LLM-powered agents utilizing autonomous web access, memory modules, and retrieval-augmented generation (RAG) are vulnerable to indirect prompt injection and environmental manipulation. Attackers can embed malicious instructions into external data sources trusted by the agent (such as Reddit posts, public databases, or ArXiv papers). When the agent autonomously retrieves and processes this content during task execution, it executes the embedded malicious commands. This vulnerability…

A vulnerability exists in large language models (LLMs) where insufficient sanitization of system prompts allows attackers to extract sensitive information embedded within those prompts. Attackers can use an agentic approach, employing multiple interacting LLMs (as demonstrated in the referenced research), to iteratively refine prompts and elicit confidential data from the target LLM's responses. The vulnerability is exacerbated by the LLM's ability to infer context from seemingly innocuous…

Large Language Model (LLM) watermarking schemes based on n-gram probability biases (specifically KGW, SynthID-Text, MinHash, and SkipHash) are vulnerable to adversarial removal during Knowledge Distillation. When a student model is trained on the output of a watermarked teacher model, it inherits the watermark's statistical biases ("radioactivity"). An attacker can exploit this inheritance by comparing the student model's output token probabilities against a base model to extract the…

Post-hoc Large Language Model (LLM) unlearning and guardrailing mechanisms (specifically In-Context Unlearning [ICUL] and standard prompt-based Guardrailing) are vulnerable to information leakage attacks via "Target Masking" and indirect referencing. These systems rely on superficial semantic matching to suppress "forget sets" (specific entities or concepts). Attackers can bypass these restrictions by querying associated properties, relationships, or pseudonyms rather than the explicit target…

Large Language Models (LLMs) used in hate speech detection systems are vulnerable to adversarial attacks and model stealing, resulting in evasion of hate speech detection. Adversarial attacks modify hate speech text to evade detection, while model stealing creates surrogate models that mimic the target system's behavior.