Safety Vulnerabilities

Large Language Models (LLMs) employing alignment techniques remain vulnerable to "jailbreak" attacks. The AutoDAN technique automatically generates semantically meaningful prompts that bypass safety features and elicit malicious outputs from aligned LLMs, unlike previous methods producing nonsensical prompts easily detectable by perplexity checks. These prompts exploit weaknesses in the LLM's alignment, causing it to generate responses that violate intended safety constraints.

Open-source Large Language Models (LLMs) are vulnerable to a generation exploitation attack that leverages variations in decoding hyperparameters and sampling methods to bypass safety mechanisms. Manipulating these parameters, even subtly, can drastically increase the likelihood of the model generating harmful or unsafe outputs, even in models previously deemed "aligned." The attack is effective even when removing only the system prompt.

Fine-tuning aligned Large Language Models (LLMs) on a small number of adversarially crafted examples, or even on benign datasets, can compromise their safety alignment, leading to the generation of harmful or inappropriate content. This vulnerability exploits the few-shot learning capabilities of LLMs, allowing attackers to override existing safety mechanisms with minimal effort and cost. Even unintentional fine-tuning with seemingly benign datasets can result in unintended safety degradation.

Large Language Models (LLMs) are vulnerable to Compositional Instruction Attacks (CIA), where malicious prompts are embedded within seemingly harmless instructions. This allows attackers to bypass safety mechanisms and elicit harmful responses from the model, even if the individual components of the prompt would be flagged as safe. The attack exploits the model's inability to correctly identify underlying malicious intent within composite instructions.

Large Language Models (LLMs) are vulnerable to In-Context Attacks (ICA) and susceptible to mitigation via In-Context Defense (ICD). ICA leverages a small number of harmful demonstration examples within a prompt to elicit harmful responses from the LLM, even if it is otherwise safety-aligned. ICD counteracts ICA by prepending safe demonstration examples to the prompt, effectively reducing the likelihood of harmful output. The effectiveness of both ICA and ICD is demonstrated across multiple LLMs.

A vulnerability in large language models (LLMs) allows attackers to craft malicious prompts that induce the LLM to generate harmful content, such as fraudulent material, racist remarks, or instructions for illegal activities. The vulnerability arises from the LLM's inability to reliably distinguish between benign and malicious instructions disguised within seemingly innocuous prompts. Attackers can exploit this by leveraging techniques like obfuscation, code injection/payload splitting, and…

Low-rank adaptation (LoRA) fine-tuning allows efficient circumvention of safety training in large language models (LLMs), such as Llama 2-Chat 70B, resulting in significantly reduced refusal rates for harmful prompts while maintaining general performance capabilities. Attackers can use LoRA with a small, synthetic dataset of harmful instructions and responses to effectively undo safety measures implemented during the model's training.

Large Language Models (LLMs), such as GPT-4, exhibit a cross-lingual vulnerability in their safety mechanisms. Translating unsafe English prompts into low-resource languages, using readily available translation APIs like Google Translate, bypasses the LLM's safety filters and elicits harmful responses with a significantly higher success rate than attacks targeting the English language directly. The vulnerability stems from an unequal distribution of safety training data across languages…

Large Language Models (LLMs) employing Reinforcement Learning from Human Feedback (RLHF) and instruction tuning methods may exhibit superficial safety guardrails vulnerable to parametric red-teaming attacks. Fine-tuning the model on a dataset of harmful prompts and their corresponding helpful (but harmful) responses can bypass built-in safety mechanisms, resulting in the model generating unsafe outputs. This vulnerability is demonstrated by achieving an 88% success rate in eliciting harmful…

A prompt-based adversarial attack, termed PromptAttack, can cause Large Language Models (LLMs) to generate incorrect outputs by manipulating the input prompt. PromptAttack crafts prompts that include the original input, an attack objective (to generate semantically similar but misclassified output), and attack guidance with instructions for character, word, or sentence-level perturbations. This allows an attacker to manipulate an LLM's response without direct access to its internal parameters…