Language Model Security Database

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…

Large Language Models (LLMs) are vulnerable to one-shot steering vector optimization attacks. By applying gradient descent to a single training example, an attacker can generate steering vectors that induce or suppress specific behaviors across multiple inputs, even those unseen during the optimization process. This allows malicious actors to manipulate the model's output in a generalized way, bypassing safety mechanisms designed to prevent harmful responses.

Standard Large Language Model (LLM) unlearning techniques, specifically Negative Preference Optimization (NPO), Gradient Difference (GradDiff), and Representation Misdirection for Unlearning (RMU), fail to sufficiently flatten the loss landscape surrounding the "forgotten" weights. This sharp loss landscape allows for a "Relearning Attack," wherein an attacker can fully restore the unlearned capabilities (such as hazardous knowledge, sensitive data, or copyrighted material) by performing…

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…

State-of-the-art machine unlearning and safety fine-tuning methods for Large Language Models (LLMs) fail to robustly remove hazardous capabilities or refusal mechanisms from model weights. While these methods suppress model outputs during standard input-output interactions, the underlying capabilities remain latent in the parameter space. An attacker with access to model weights (e.g., via open releases or leaked weights) can restore "unlearned" knowledge (such as dual-use biology hazards) or…

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.

Large Language Models (LLMs) employing alignment techniques for safety embed a "safety classifier" within their architecture. This classifier, responsible for determining whether an input is safe or unsafe, can be approximated by extracting a surrogate classifier from a subset of the LLM's architecture. Attackers can leverage this surrogate classifier to more effectively craft adversarial inputs (jailbreaks) that bypass the LLM's intended safety mechanisms. The attack success rate against the…

Large Language Models (LLMs) are vulnerable to jailbreaking attacks that manipulate attention scores to redirect the model's focus away from safety protocols. The AttnGCG attack method increases the attention score on adversarial suffixes within the input prompt, causing the model to prioritize the malicious content over safety guidelines, leading to the generation of harmful outputs.

A Cross-Prompt Injection Attack (XPIA) can be amplified by appending a Greedy Coordinate Gradient (GCG) suffix to the malicious injection. This increases the likelihood that a Large Language Model (LLM) will execute the injected instruction, even in the presence of a user's primary instruction, leading to data exfiltration. The success rate of the attack depends on the LLM's complexity; medium-complexity models show increased vulnerability.

Large Language Models (LLMs) are vulnerable to jailbreaking attacks leveraging synthetically generated prompts. A novel pipeline, SAGE-RT, generates a diverse dataset of 51,000 prompt-response pairs designed to exploit LLMs' vulnerabilities across various categories of harmfulness. These prompts successfully jailbreak state-of-the-art LLMs in a significant percentage of tested sub-categories, including 100% of macro-categories for certain models like GPT-4 and GPT-3.5-turbo. The vulnerability…